Skip to main content

# Full text of "Equilibrium distribution of zeros of random polynomials"

## See other formats

```EQUILIBRIUM DISTRIBUTION OF ZEROS OF RANDOM

POLYNOMIALS

BERNARD SHIFFMAN AND STEVE ZELDITCH

Abstract. We consider ensembles of random polynomials of the form p(z) = X^=i a jPj
where {oj} are independent complex normal random variables and where {Pj} are the
orthonormal polynomials on the boundary of a bounded simply connected analytic plane
domain fl C C relative to an analytic weight . In the simplest case where ft is the

unit disk and p — 1, so that Pj{z) — z J , it is known that the average distribution of zeros
is the uniform measure on S 1 . We show that for any analytic (f2, p), the zeros of random
polynomials almost surely become equidistributed relative to the equilibrium measure on
d£l as N — > oo. We further show that on the length scale of l/N, the correlations have a
universal scaling limit independent of (£l,p).

Introduction

A well-known result due to Hammersley [jHam|| (see also Shepp-Vanderbei ||S V|| ) states
that the zeros of random complex 'Kac' polynomials

N
3=0

tend to concentrate on the unit circle S 1 = {\z\ = 1} as N ^ oo when the coefficients cij are
independent complex Gaussian random variables of mean and variance 1:

E(tXj-) = 0, E(a 3 -ajfc) = 5 jk , E(a i a fe ) = 0. (2)

To be precise, (0) defines a Gaussian probability measure 7^ on the space Vn of polynomials
of degree < N on C, studied long ago by Littlewood-Offord, Erdos-Turan, and in particular
by Kac [ |KT| , K2|| . We denote expected values relative to 7^ by E^i. We also define the
normalized distribution of zeros of / to be the probability measure

^ == ^ E *• • (3)

/W=o

Then we may formulate the concentration of zeros on S 1 as the weak limit formula:

E « (Zf) ^5 s i as N ^ 00 , (6si,<p) := ^ / v(e w ) d9. (4)

The first purpose of this note is to generalize the expected equidistribution result (||) in a
suitable sense to any closed analytic curve dQ in C which bounds a simply connected plane

Date: June 17, 2002.

Research partially supported by NSF grants #DMS-0100474 (first author) and #DMS-0071358 (second
author) .

2

BERNARD SHIFFMAN AND STEVE ZELDITCH

domain Q. Our method is new even in the case of S 1 and gives sharper results than those of

Hamf . We further prove that the equidistribution result is self-averaging in the sense that
Zf N is almost surely asymptotic to the expected distribution. Our second purpose is to show
that the correlations between zeros have a universal scaling limit independent of Q, p on the
length scale of 1/N and to determine the properties of the limit pair correlation function.

Our starting point is to reinterpret the Gaussian measure (0) on Vn as the one induced
by the inner product

1

2n

(f'9) s i = — I f9\dz\ (5)

on Vn- In general, any inner product

(f>9)u = / fddfi
Jc

on Vn induces a Gaussian measure on Vn as follows: Let {Pj} denote an orthonormal basis
of Vn relative to (, ) and write any polynomial in this basis:

N

K-) (6)

3=0

The Gaussian measure is defined by the condition that the coefficients aj in this basis are
i.i.d. complex Gaussian variables with mean zero and variance one; i.e. it equals n~ N e~^ 2 da
in terms of the coefficients.

The choice of the Gaussian measure 7^ in @ results in the zeros of the random poly-
nomials becoming equidistributed relative to 5gi. As is well-known, 8\$i is the equilibrium
measure of S 1 (or of the closed unit disk). Recall that the equilibrium measure of a compact
set K is the unique probability measure Ujc which minimizes the energy

E(p) = — / / log \z — w\ dfi(z) dfi(w).
Jk Jk

We claim that the same equidistribution result holds if we replace the unit disk U = {\z\ < 1}
by any simply connected, bounded plane domain Q with real-analytic boundary. That is,
we replace the inner product (El) on S 1 with an inner product on dQ of the form

(/, g)an, P := / f(z)g(z) P (z) \dz\ , p g c-(an) . (7)

Jan

We denote by JoQ tP the Gaussian measure induced by this inner product on Vn as in @,
and we denote the expectation relative to the ensemble (Vn, ldn )P ) by . As in the
S 1 case, we obtain an asymptotic formula for the expected normalized distribution of zeros

Theorem 1. Suppose that Q is a simply- connected bounded C u domain and p is a positive
C w density on dfl. Then with the above notation,

*&JZ?)= Va + 0{\/N) ,
where uq is the equilibrium measure ofCl.

EQUILIBRIUM DISTRIBUTION OF ZEROS OF RANDOM POLYNOMIALS

3

Here we use the expression 0(1/N) to mean a distribution un £ P'(C) such that | (un, ¥>) | <
C v /N for all ip G 22(C) , where C v independent of N. Note that the limit distribution of
zeros concentrates on dfl = Supp (uq) and is independent of the density p used to define the
inner product.

The equilibrium measure may be understood as the measure by which N 'electric charges'
in Q distribute themselves in the limit iV — > oo. Thus, at least on average, the zeros of
random polynomials in the ensemble (Vn, 7an p ) behave like electric charges. This suggests
that the same equidistribution result should hold if we orthonormalize the polynomials within
Q. We restrict to Lebesgue measure in Q and define the inner product

(f,g)n-= / f(z)g(z)dxdy (z = x + iy)

N ■

which induces the Gaussian measure on V

Theorem 2. Suppose that Q is a simply- connected bounded domain and let denote
the expectation with respect to the Gaussian measure 7q on Vn- Then

ES(zy,<p)=»a + 0{l/N) .

As in |[ShZl|| , the expectation results of Theorems [TJ can be improved to give an almost
sure limit distribution of zeros of random sequences of polynomials of increasing degree:

Theorem 3. Suppose that Q is a simply- connected bounded C w domain. Let 'Jn be the

Gaussian measure on Vn induced by the inner product (j^) or and let fi := Yl^=i In be

the product probability measure on S := n^=i^iv- Then Z^ N — > in the measure sense,
for p-almost all sequences {Jn} £ <S; i-e.,

— <p(z) — > / (pduQ Wip G C(C) , p-almost surely .

{zeC:f N (z)=0} JdU

The principal ingredients in the proofs of Theorems [I] and |2| are classical results of Szego
|Szl| and of Carleman |Ca| ] on the asymptotics of orthogonal polynomials normalized on the

boundary and on the interior of a domain, respectively. We will recall these results in §|T[
The proofs of the two theorems are then essentially the same.

Theorems [I|-|3| say that in the limit the zeros tend to be uniformly distributed along dfl
(with respect to the equilibrium measure). However, as in our earlier work with P. Bleher
||BSZ1| , [BSZ2 1, the zeros do not behave as if they were thrown down at random; indeed, the

zeros are correlated. To quantify this correlation, we consider as in ||BSZ1| , |BSZ2|| the t-point
correlation functions

t

Kt£Jz 1 ,...,zt):= L -^, (zi,...,z t ) G C t , (9)

where

Q = {(21, . . . , zt) G C e : zj ^ z k for j ^ k} .

(We punch out the big diagonal in C since the numerator in (|9]) has a singular part there.)
The correlation functions Kq(z,w) are similarly defined.

4

BERNARD SHIFFMAN AND STEVE ZELDITCH

In particular, the pair correlation function Kq\$ (z, w) (or Kq N (z, w)) can be interpreted as
the conditional probability of finding a zero at w given that there is a zero at z, normalized by
dividing by the unconditional probability of finding a zero at w. For example, if Kq^ p (z, w) =
1, then the existence of a zero at z has no influence on the probability of finding a zero at w.
This is always the case in the limit as iV — > oo if z, w are fixed. However, we have nontrivial
correlations if the distance between points is 0(1 /N). To describe this phenomenon, let
Zq G dQ be fixed, and choose the complex coordinate £ = <&(z), where \$ is the Riemann
mapping function mapping the exterior of Q to the exterior of the unit disk, mapping z to
1, and taking oo to itself. (This exterior Riemann mapping function is a basic ingredient in
the Szego and Carleman asymptotics used in this paper.) We then have universal scaling
limit zero correlation functions:

Theorem 4. There exist universal functions K £o ° : Ce — > R + independent ofQ,zo,p such
that

*£, (i + £,..., i + 0)

as N — ► oo ; where Kqq p = o is the correlation function written in terms of the

complex coordinate ( described above. Similarly, (l + ^, . . . , 1 + ^) — > K ico ((i, . . . , Q).

The universal scaling limit pair correlation function K 2o ° is given by formula (|33|).

In our work with P. Bleher |[BSZ1| , |BSZ2|| , we showed that the zero correlation functions

can be given by general (universal) formulas involving only the Szego kernel and its first and
second derivatives. In §|3|, we study the partial Szego kernel

N

S N (z,w) ^p^)nR , (10)

k=0

which gives the orthogonal projection onto the span of the first iV orthogonal polynomials
associated to the inner product (|7]). We show (Proposition [TTJ) that the scaling asymptotics
of Sn have the form

(l + |, 1 + |) - C n , P , Z0 G(Ci + 6)

where = S^- o \$ 1 and

G(z) = — . (11)

z

It is natural to separately consider the 'tangential' scaling asymptotics along the boundary
and the 'normal' scaling asymptotics orthogonal to the boundary. In the tangential case,
we set Ci = i0, C2 = and obtain the tangential scaling limit kernel G(e %e ) = e %e / 2 ( ^jt^- ) ■
This modified sine-kernel is reminiscent of the scaling asymptotics of the projection kernels
/Ctv for the orthogonal polynomials which occur in the theory of random matrices (see ||Dc

Theorem 8.16). However, our actual result is somewhat different and the methods have little
in common.

The tangential scaling limit for pair correlations between zeros,

K T (a) := K^(0,ta) = ^K^ p (1, 1 + §) = g^KSi {^ ia ' N ) ,

EQUILIBRIUM DISTRIBUTION OF ZEROS OF RANDOM POLYNOMIALS

5

measures the probability (density) of finding a pair of zeros in small disks around two points
on dfl in terms of the (-^-scaled) angular distance a between them. As is illustrated by the
graph of k t (drawn using Maple™) in Figure [I], i n tangential directions the zeros 'repel'
as a — > and are oscillatory with 1/a 2 decay over long scaled distances, very much as for
correlations between eigenvalues of Gaussian random Hermitian matrices.

Figure 1. The tangential pair correlation function k t [

a)

In the normal directions, we obtain the scaling limit:

K x (r) := K 2oo (0,t) = lim K (

2N

T

The graph of k in Figure is reminiscent of the Bleher-Di correlation function |[BD|| for
real Kac polynomials. Although the normal correlation is not oscillatory, we again have zero
repulsion and 1/r 2 decay.

FIGURE 2. The normal pair correlation function K ± (r)

6

BERNARD SHIFFMAN AND STEVE ZELDITCH

We end the introduction with some remarks on and comparison to prior results in this
area. Real Kac polynomials, i.e. polynomials as in (|l|) but with real Gaussian coefficients of
mean zero and variance one, were shown by Kac | KT| , |K2|| to have on average around log N
real zeros. The same was proved for more general Kac polynomials in \^M\. As mentioned
above, Hammersley |[Ham|| obtained the concentration of zeros of complex Kac polynomials
on the unit circle in a somewhat less precise sense than the one we state here. Later, Shepp
and Vanderbei [ 5V | showed that the complex zeros of random real Kac polynomials also
tend to concentrate on the unit circle, explaining in a qualitative way why so few zeros are
real. Other classical studies of random real polynomials are given in ||EO| , |LO| , |E1| , |EKf .
The probability of a real Kac polynomial having no real zeros is estimated in [DFSZJ. Other
results on hole probabilities for complex zeros are given in ]So|.

Recently, a number of results were obtained on the distribution of and correlations be-
tween zeros of Gaussian random polynomials in one variable (and more general holomorphic
sections) in which the inner product comes from a hermitian metric h on a positive line
bundle L — > M over a Riemann surface M (see ||Han| , BR , |ShZl| , [BSZ1|| . For simplicity, let
us mention only the case of SU(2) polynomials, in which M = CP X ,L = O(N), and where
h = hp S is the N th power of the Fubini-Study metric. It is obvious from the SU(2) symmetry
of the expected distribution of zeros in this case that it must equal the standard (Fubini-
Study) area form, in sharp contrast to the very singular equidistribution result we find in
Theorems [l|-|3]. Moreover, although the pair correlation function for SU(2) polynomials given
by ||Han|| (see [[BSZ1|] for a generalization) exhibits the same repulsive behavior near 0, its

scaling limit occurs on the length scale of 1/yN rather than 1/N. Thus, the correlations are
quite incomparable in the two types of ensembles.

To explain the strong differences in the results, it should be observed that the ensemble
(Vn, In) in the SU(2) case comes from an inner product {, ) N on Vn which changes with each
N, according to which ||z fe ||jv = (T). These inner products (,)jv derive from a single inner
product on CR (Cauchy-Riemann) functions on the boundary of the unit ball in C 2 rather
than on polynomials in one complex variable. This suggests that polynomials in the SU(2)
ensemble should be thought of as homogeneous polynomials Y2j=i a j z o ~" 1 z i * n ^ wo variables
and that their equidistribution law and correlations reflect the extra dimensionality. Because
of the concentration of zeros on curves, the ensembles studied here share a one -dimensionality
with the ensembles of GUE and CUE random matrices and with the observed correlations
between zeros of the Riemann zeta function.

We note that in random matrix theory, orthogonal polynomials on R with respect to
weights pN = e~ NV ( xS> dx depending on the degree N are quite important (see ||De| | ) . This
suggests that ensembles (Vn,1n) of random polynomials on curves dfl where the inner
product is given by a weight of the form Pn = p(z) N \dz\ might be interesting. Such ensembles
have features in common with both the SU(2) ensemble (and its generalization to other
Riemann surfaces) and with the ensembles of this paper. How are the zeros distributed and
correlated? Do the zeros concentrate on dVt or are they diffuse? Our methods do not seem to
adapt in a simple way to such ensembles, and perhaps they belong to different universality
classes than the ones studied here.

Another interesting direction would be to generalize our results to higher dimensions.
Clearly the methods we use, based on the Riemann mapping theorem, have no simple gen-
eralization.

EQUILIBRIUM DISTRIBUTION OF ZEROS OF RANDOM POLYNOMIALS

7

1. ASYMPTOTICS OF ORTHOGONAL POLYNOMIALS

The proofs of Theorems [L|-|2| are based on asymptotic properties of orthogonal polynomials
associated to fl, and to their relations with the interior and exterior Riemann mapping
functions. These imply asymptotic properties of partial Szego and Bergman kernels for fl.
We recall the results we need in this section. Classical references are [ 5zl , |Ca|| while more
recent references (with more precise results) are in [^1], pu |.

1.1. Szego kernel and orthogonal polynomials. Let fl C C be a smooth bounded
domain. The Szego kernel of fl with respect to a measure p\dz\ on dfl is the orthogonal
projection

S : £ 2 {dfl, p\dz\) -> H 2 {dfl, p\dz\) (12)

onto the Hardy space of boundary values of holomorphic functions in fl which belong to
C 2 (dfl, \dz\). The Schwartz kernel of S is denoted S(z,w). According to |[Bel|| , Theorem
24.3, S(z, w) admits an analytic continuation (holomorphic in z and anti-holomorphic in to)
to fl x fl \ A, where A is the diagonal in dfl x dfl. We will refer to this as the regularity
theorem.
Let

{Pj(z) = a j0 + cijiz H h ajjZ j }

be the orthonormal basis of orthogonal polynomials for C 2 (dfl, p\dz\) obtained by applying
Gram-Schmidt to {1, z, z 2 , . . . , z\ . . .}. Since S is an orthogonal projection, we may express
it in terms of any orthonormal basis. Hence, we have

oo

S(z,w) = ^2P k (z)P k {w), {z,w)eQxQ (13)

k=0

By the regularity theorem, one has that S(z, z) < oo for z G Q, and thus P/v ~~ ► on Q.
Hence,

Sn(z, z) — > S(z, z), uniformly on compact subsets of fl,

where S^(z,w) = J2k=o Pk{z)Pk{w) is the partial Szego kernel given in fliPp.
In the case of the unit disk U (with the weight p = 1), one has:

S u (z,w)

2tt(1 — zw)

We observe that S(z, w) admits a meromorphic continuation to the exterior as well as the
interior of Q. However, the behavior is quite different with regard to the partial Szego kernels.
Indeed, The orthogonal polynomials are Pk{z) = z h , hence

N-l , _ I \2N

\p N {z)\ 2 = \z\ 2 \ ®M) = EM tt = T4r- ( 14 )

fc=0 ' '

Clearly, Sn(z,z) —>■ oo at an exponential rate in the exterior of fl.

BERNARD SHIFFMAN AND STEVE ZELDITCH

1.2. Partial Szego kernels and the exterior Riemann mapping function. We will

need the behavior of Sn{z, z) in the exterior of Q. The relevant result was proved by Szego

Let f2 C C denote a simply connected bounded plane domain with C°° boundary <9f2. The
exterior domain C \ Q is also a simply connected domain in C and we denote by

\$:C\fi^C\£7, = cz + c + c x z~ x + ■■■ (15)

the (unique) exterior Riemann mapping function with \$(oo) = oo, \$'(oo) G M + . We recall
that the equilibrium measure uq of Q is given by

^ = \$*<5 S i; i.e., J<pdv n = -^j\o<f>- 1 (e ie )dB. (16)

Remark: If we fix a point ^ 6 fl, we may define as well the interior Riemann mapping
function / : Q — > U with f(xo) = 0, /'(%) G M + . The functions f\gn and are generally
different.

Let us assume that dQ G One can define a (unique) nonvanishing holomorphic function
A e on C \ Q satisfying:

l A elL = P ■

(See [^z|, §§10.2,16.1].) In the following, L = Length(dfi), and

T e (dn) = {x G C : d(x, dVl) < e}.

Theorem 5. |(5z2| , Theorems 16.4-16.5] Lei denote the orthonormal polynomials rel-

ative to (dfl, p\dz\), where Q, p are as in Theorem^. Normalize P n so that its leading coef-
ficient is positive. Then there exist e > and < 5 < 1 such that for z G (C \ Q) U T e (dQ),
we have

1/2

Pn(z) = (J^J A^)- 1 ^) 1 / 2 ^)" + 0{5 r '

The proof is based on the properties of Faber polynomials F n associated to G n (z) : =
{^) 1/2 A e {zY^\z) l l^{zY- i.e., F n (z) is the polynomial part of the Laurent expansion of
G n {z) about z = oo. Suppose that the Riemann mapping function \$ _1 extends analytically
to {\z\ > r} where r < 1. The main estimates are:

P n = F n + 0(n l ' 2 r n l 2 ) = G n + Oin 1 ' 2 ^' 2 ) .

This shows that the Faber polynomials (with weights) are asymptotic to the orthogonal
polynomials (with weights).

1.3. Bergman kernel and orthogonal polynomials. In the proof of Theorem |2j on the
expected distribution of zeros, where the inner product is chosen to be (/, g)n = Jq fg dx dy,
the role of the Szego kernel is played by the Bergman kernel, i.e. the orthogonal projection
from £ 2 (Q) onto the subspace 7i 2 (Q) spanned by the C 2 holomorphic functions.

As before, we assume Q C C is a simply connected, bounded plane domain with C w
boundary. We denote by

{Pj(z) = a j0 + a n z + ■■■ + a jjZ J }

EQUILIBRIUM DISTRIBUTION OF ZEROS OF RANDOM POLYNOMIALS 9

the orthonormal basis of orthogonal polynomials for £ 2 (Q,dxdy) with positive leading co-
efficient. The Bergman kernel may be expressed in terms of the orthogonal polynomials
by:

oo

B{z,w) = ^P k {z)P k {w) , (z,w)eQxQ. (17)

fc=0

We let B>n(z, w) = J2 k=0 Pk(z)Pk(w); as in the case of the Szego kernel, we have

B n (z, w) — > B(z, w) , B(z,z)>0.

Orthogonal polynomials over a domain were studied by Carleman after Szego's work on
polynomials orthogonal on dfl. A key element in our proof is the following asymptotic
formula for P/v(z):

Theorem 6. (Carleman [Pap ; cf. |fSL| , Theorems 1-2, p. 290) Let {P n } denote the orthonor-
mal polynomials relative to (Q,dxdy), where Q is a simply- connected bounded C u domain.
Normalize P n so that its leading coefficient is positive. Then there exist e > and < 5 < 1
such that:

{ (^) /2 <S>'(z)\$(z) n [l + 0(5 n )}, zeC\Q

2. Proof of the asymptotic formulas

We now prove Theorems |T]-[^. The proofs are basically the same in the boundary and
interior cases, granted the theorems of Szego and Carleman. We therefore only give details
in the boundary case.

We first relate zero distributions to Szego kernels, as in our previous papers R5hZl| , |ShZ2
2.1. Zero distributions. For convenience, we let

7iV

denote the zero distribution of a polynomial / so that recalling (|^), we have Z
(Strictly speaking, we should count the zeros with multiplicities, but polynomials in Vn
almost surely have only simple zeros.) The expected zero distribution has a simple expression:

Proposition 7. We have

E m, P ( Z f">

dd log Sn(z, z).

This proposition is a special case of Proposition 4.1 in | |ShZ2| | (which is a variation of a
result in our earlier work |[5hZl|| ). For completeness, we give the proof here: We first note
that since

7 log m 2 ,

2tt

10 BERNARD SHIFFMAN AND STEVE ZELDITCH

we have

EX(^/) = ^^ E ip(log|/| 2 ) •
To calculate the expectation, we write / in terms of the orthonormal basis {Pj} of 7-V:

N
3=0

where a = (ao, . . . , ajv), P = (Po, • • • , -Pat). Then,

^m, P (Z f ) = ^dd [ log | (a, P(z))\ -4ne-^ 2 da.

We write

N

P(z) = \\P(z)\\u(z), \\P(z)\\ 2 = Y,\PAz)\ 2 = Sn(z,z), |K«)|| = 1.

3=0

Then,

log | (a, P(z))\ = log \\P(z) || + log | (a, u{z))\ .

We observe that

/ log \{a,u(z))\ e " a " 2 rfa = constant

ic JV + 1

since for each 2; we may apply a unitary coordinate change so that u(z) = (1, 0, . . . , 0). Hence
the derivative equals zero, and we have

E^ p (Z f ) = y-^dd\og\\P(z)\\ = \±ddlogS N (z,z) .

□

By exactly the same argument, we also have:
Proposition 8. We have

K{Zf) = ^dd\ogB N {z,z).

2.2. The circular ensemble. We start with the fundamental case of the circle; in the next
section we will reduce the other cases to the circular case. Before proving our precise result
for the circle (Proposition we first give a simple ad hoc argument that the expected limit
distribution of zeros in this case is the measure 5gi given in @): By ([14]),

^dd\ogS N {z,z) ~ ia§log(l - \z\ 2N ).

Clearly, in any annulus \z\ < r < 1, (1 — |^| 2JV ) — > 1 rapidly with its derivatives, and the
limit equals zero. In any annulus \z\ > r > 1 we may write (1 — |2| 27V ) = |z| 27V (|z| _2Ar — 1)
and separate the factors after taking log. The second again tends to zero rapidly, while the
first factor, log \z\ 2N , is killed by dd (note that z 7^ in this part). It follows that the limit
measure must be supported on S 1 . Since it is SO(2)-invariant (radial), and since it is a
probability measure, it must be j^d9, which we henceforth denote by v.

We have the following explicit formula and asymptotics for the circular case:

EQUILIBRIUM DISTRIBUTION OF ZEROS OF RANDOM POLYNOMIALS

11

Proposition 9. Let v — ^ denote Haar measure on S 1 . Then

21 ~\2N

UN

(N+l) 2 \z
(\z\ 2N + 2 - l) 2

~2n~

dz A dz

Furthermore, E;; {Zf) = Nu + 0(1); i.e., for all test forms <p G T>(C), we have

E

N

,WW=o}

/■27T

/ <p(e ie ) ^ + 0(1) .

2n~

In particular, E^ (Zj^) v in V(C).

The formula for E^ (Zj) in Theorem |9| agrees with the one given by Hammersley ||Ham |
(see also fEK],[SV|).

Proof. By Proposition |7|,

K(Zf)

2n

-dd log

1 -

z

2N+2

1 -

Z

2

We write ( = p + i9 = 2 log z, and we let

rr 1 i A l_ e (iv+i)p

n=0

1 -

Then

We let

^(Z f ) = —{\ogh N )"(p)d(d(.

in

g N : = (log/ijv)"

(iV+l) :

-(JV+l)p

(19)

(20)
(21)

(1 -e-p) 2 v " ' (1 - e-( Ar + 1 )^) 2 '

The desired formula for E^ (Zf) follows by substituting d( = 2 z dz into (PH|)-([2T|).

We easily see that > (which also follows from the strict subharmonicity of
\ogS%(z,z) = log(£^ =0 M 2 ™)) and that 9n(~p) = 9n(p)- Now suppose that ip e C°°(R) n
£°°(M) such that ^(0) = 0. We claim that

9Ni> = 0(1)

(22)

To verify Q22fl, we break up the integral into 3 pieces: p < — 1, — 1 < p < 1, and p > 1.
First, we have

9n

+oo

9N

1

N+l 1

and thus

-i

9Nip

e - 1 e^ 1 - 1 ~ e - 1

+oo

9Nip

0(1).

To estimate the integral over [—1, 1], we write ip(p) = ap + 0(p 2 ) and we note that
9n(p) <

1 4

< — for < p < 1

(1 - e-p) 2 ~ p 2

12 BERNARD SHIFFMAN AND STEVE ZELDITCH

Since qn is even, we then have

9 Nip

<C / P 2 g N (p)dp<4C

which completes the proof of fl22|) .

Now let cp G V{R). An easy computation gives j R g^dp = N. (This also follows from the
fact that the expected number of zeros of a degree- N polynomial is N.) Thus by

9NV = <p(0) 9n+ g N [<P - <p(0)] = N<p(0) + 0(1)

□

Remark: Recalling (jl9|) , we have

^-1(0)^(0) - ^_ x (0) 2 _ NJ2to k 2 ~ (EiS kf

9N-i(0)

N 2

N 2 -l
12

One can easily check that g' N (p) < for p > and since is even, we have <7jv(0) = supt/jv-
In fact, a computation using Maple™ gives:

N 2

N 4

N 6

9N-1

12

240

6048

1 4 iV 8 -l^ 6+ A~

10

•P

691(iV

12

172800'

5322240'

118879488000

10

2.3. The general case. We now prove Theorem [I]. We must show that

2tt

-ddlog S N (z,z) = Nu n + 0(1)

(23)

We shall break up the proof of (Effi) into two regions: fl and a neighborhood W £ of C \ fl.
The estimate (p3|) is obvious on f2. Indeed, by Theorem 16.3 in ||Sz2|| , Sn{z,z) — ► S{z,z)
uniformly in compact subsets of fl Furthermore, S^z, z) > on fl (see e.g., |[Bel| , Th. 12.3]),
and hence log Sn (z, z) — > logS(z, z). Thus

dd\ogS N (z,z) = dd\ogS(z,z)+o(l) = 0(l) in V'(fl) .

(In fact, Sn(z,w) — > S(z,w) on fl x fl, and hence by the Cauchy Integral formula, we have
normal convergence of all derivatives; it then follows that 3d log Sn(z, z) — > 99 log S^z, z)
uniformly on compact subsets of fl.)
Next we verify (^3|) on the domain

iy £ :=(C\fi)UT £ (9fi),

where e is chosen sufficiently small so that Theorem [5] holds and \Q\ > 5' > 5 on W £ .
Let

A N (z) = S N (z,z)/<S>* N S^(z } z) .
We claim that there is a positive constant C such that

< 1/C < A N (z) < C < +oo for z G W £

(24)
(25)

EQUILIBRIUM DISTRIBUTION OF ZEROS OF RANDOM POLYNOMIALS 13

T \ 1/2

1>{z) = ( — KizY^'iz) 1 ' 2 , (26)

To verify (25), we let

and we note that ip an d \ are bounded on W e . Recalling that

N

S^(*,z)=][>(z)|

2n

n=0

it then follows from Szego's Theorem || that sup z&We A^(z) < C < +oo.

The lower bound for A^(z) follows from [5z2 , §16.5]. Alternately, let G n (z) := ip(z)§(z) n
as above. Then by Theorem |, there exists n > 2 such that |P n (;z)| 2 > ||G n (z)| 2 for
zeW £ , n> no. If 5' < ||\$(z)| < 1 we have \p (z)\ 2 = c> c'(\G (z)\ 2 + • • • + \G no (z)\ 2 ) and
hence

N

S N (z,z) > min{l/2,c'}]r|G n ^)| 2 > d'^ N S u N {z,z) . (27)

n=0

On the other hand, if |\$(^)| > 1 then the required estimate follows from

Sn(z,z) > \P no (z)\ 2 + ■ ■ ■ + \P N (z)\ 2 > l[\G no (z)\ 2 + ... + \G N (z)\ 2 ]
> \[\G (z)\ 2 + ... + \G N (z)\ 2 ] =\\^z)\ 2 ^ N S u N {z,z),

for N >2n .

It follows from (|25|) that

[ddlog A N ,<p) = (log A N ,dd<p) = 0(1) , for if e V(W £ ) .

Recalling ([IB]), we then have

^-ddlog S N (z,z) = \$*(^dd\ogS%(z,z))-^ddlogA N (z)

<f>*(Nv + 0(l)) + 0(l) = Nu n + 0(1)

□

2.4. The interior case. The proof of Theorem |2| is exactly the same as in the boundary
case, using Carleman's theorem in place of Szego's, and the Bergman kernel and Proposition |8]
in place of the Szego kernel and Proposition [7|. We omit the details.

We also note that the conclusion of Theorem [| holds for inner products on Q with certain
analytic weights; see [ 5I| .

2.5. Proof of Theorem ^. The proof of almost sure convergence to the average is exactly
the same as the proof of the similar statement for sections of positive line bundles in | |SliZl
We summarize the proof here: Let

UJ N

-^dd log Sn(z, z) for the boundary case,
■^fdd log -Bat (z, z) for the interior case.

14 BERNARD SHIFFMAN AND STEVE ZELDITCH

We then have the following variance estimate:
Lemma 10. Let if be any smooth test form. Then

E^Zy-UN,^*) =0(N- 2 ).
Lemma [10] is given as Lemma 3.3 in ||ShZl|| and has exactly the same proof. Continuing

as m

ShZl|| , by Theorem [1] or Theorem 0, it suffices to show that
(Zf N — lun, ip) — > almost surely .
Consider the random variables Xn on S given by:

X N ({f N }) = (Zf N -cu N ,ip) 2 >0.
By Lemma |TD|, f s X^ dp = 0(N~ 2 ), and therefore

» oo oo „

/ Xjy dp = / Xn dp < +oo .
Js N=1 N=1 Js

Hence, Xjy — > almost surely. The conclusion follows by considering a countable C°-dense
family of test forms. □

3. Pair correlations of zeros

In our previous work with P. Bleher |BSZ1| , |BSZ2|1 , we derived the scaling limit pair
correlation functions for zeros of SU(m + l)-polynomials and showed that these correlations
are 'universal.' In this section, we apply the general formulas from [[BSZ1| , [BSZ2|| to prove
Theorem and to describe our universal pair correlation function K 2o °.

3.1. The scaled Szego kernel. In [ |BSZ1| , §2.3], we gave a general formula for the £-point
correlation of zeros in terms of the Bergman-Szego kernel and its first and second derivatives.
In order to apply this formula to find the scaling limit correlations, we need the following
universal scaling limit Szego and Bergman kernels for plane domains:

Proposition 11. Let Q be a simply- connected bounded domain and let p be aC u density.
i) Let Sn be the orthogonal projection for the inner product (, )an,p- Then

\s N (l + %, 1 + mz )\ 2 G(Ci + CO , G(z) = ,

N \ N N /

where Sn = SV 0< £ _1 is the projection kernel written in terms of the complex coordinate
C = \${z), i> is given by fl2(\ ) and z = \$ _1 (1).
ii) Let Bn be the orthogonal projection for the inner product (,)n- Then

where = B^o^ 1 is the projection kernel written in terms of the complex coordinate
C = \$(z) and z = \$ _1 (1)-
The above limits are uniform when |Ci| + IC2I is bounded.

EQUILIBRIUM DISTRIBUTION OF ZEROS OF RANDOM POLYNOMIALS 15

Proof. We begin as before with the case of the disk U with density p = 1. Then
1 1 N

tX(1 + Ci/N, 1 + (z/N) = -^(l + Ci/iV) fc (l + C 2 /iV) fc

fc=0

[(i + Ci/iV)(i + c 2 /iV)]
d + G + CiG/iv

e Ci e C2 _ i

jV+1

G(Ci + Ca) • (28)

U + t.2

We now consider the inner product (, )an,p- By Szego's Theorem [5], we can choose e, A >
such that

P fe (» = ^{z)[^{z) k + g fc (z)] , \q k {z)\ < for 1 - e < \®{z)\ < 2 ,

and hence

S N {1 + 77; 1 + ) _ 1 , Cl 1 , , 1 I en* JTTjT

ATV(1 + + |) ^

i * i N

fc=0 fc=0

By (|28|) , the first term on the right side approaches G(Ci + (2)- Thus it remains to show that
the other terms tend to zero. Suppose that < A < +00. Then,

1 ±(1 + |)*STT|) < |fe«.:/-» < § + § £ = o (1) .

fc=0 fc=0 ~ k=k V 7

Exactly the same estimate holds for the next term, and the last term is clearly also 0(jt).

The same argument holds for the partial Bergman kernel, using Carleman's Theorem
instead of Szego's. □

3.2. The scaled zero density. We define the expected zero density function Dq U by

KnAZf) = D^ p (z) (idzAdz) .
We recall that by Proposition |9|,

We now give a third limit for Djj x , the scaling limit, which we find to be universal:

Proposition 12. LetVL,p,§ be as above and let D N :— (D N o \$ _1 )|(\$ _1 ) / | 2 be the expected
zero density for the inner product (, )dn, P or for (, }q with respect to the coordinate ( = \$(2).
Then

N 2

D N (l + ^ -f D°°(t) as N -> 00 ,

16

where

BERNARD SHIFFMAN AND STEVE ZELDITCH

D^{r) = -{\ogG)'\2r) = -

4t

(2 + 4r 2 )e 2T + l 1

7T- " ' ' ' An(e 2T - 1)V

vrr 2 4vr(e 2r -l) 2

In particular, the scaling limit density D°°(t) has 1/r 2 decay as r — > ±00.
Proof. By Proposition 0,

5 » = log §v(r,r).

The conclusion follows by substituting r = 1 + jr and applying Proposition [□].
The graph of the scaling limit density function is given in Figure |3| below.

□

0.026-

fl.02-

/ 0.015 -

/ 0.01

/ 0.005

Figure 3. The scaled zero density D°°(t)

3.3. The scaling limit zero correlation functions. Theorem ^ is an immediate conse-
quence of Propositions [11-12 and Theorem 2.4 in PSZ1| (with k = m

1, n — £), which

gives a universal formula for the £-point zero correlations in terms of the projection kernel
and its first and second derivatives. (See also [BSZ2| , Theorem 1.1].)

For the reader's convenience, we derive the formula for the pair correlation case £ = 2:
Using the coordinate £ = &(z), we define the 2x2 matrices

Sn((,() Sn{(,v)
S N (r},() S n (t],t])

o 2

d(dr]

As in pSZlj pSZ2| we let

Writing

A N
E N

C N - (B 1

B

m>p (z? xz i ;) = \ N ((, v )dVo\ €2

EQUILIBRIUM DISTRIBUTION OF ZEROS OF RANDOM POLYNOMIALS 17

we have by PSZIj (91)] or |B^Z2| , (23)],

A n A 22 + A^A^
Xn = 7r 2 detA^ • (29)
We consider the case of the inner product (, )dn, P - (The interior case is similar.) By

Proposition 11

£A»(1 + (l ,N, 1 + (2 /N) ( g<£ + |j g«j + |) j * A » (Cb 6) _ (30)
where c = ^(zq)! -2 - Differentiating (pO]) , we obtain

b^(i + Ci/iv, i + C2/A0 - (^(0 + a)) iift=1)2 = ■b-(c 1 ,c 2 )

^c^(i + G/iv,i + c2/A0 - (G"(o+a)) iife=lj2 = c-(Ci,c2).

We write

A°°(Ci, C2) = C°° - (B 00 )*(A 00 )- 1 B 00 = lim -^-A N (1 + (JN, 1 + ( 2 /N) . (32)
Hence by (p9l),

^ M i + {iW i + 6 /jv)^^±^.

Therefore by Proposition [12| ,

A , Ci 1 , C2 \ 7^200//- > \ _ (A^A^ + A^A^)(d, (2) /„o\

\ AT' N J Ui'W ~ Doo{Re 0) £>oo (Re Ca) det A-(Ci,C 2 ) ' 1 j

□

(31)

The explicit expansion of (|33| ) in terms of Ci, C2 is quite complicated. Using Maple™, we
can compute the expansions of the tangential and normal correlations for short distances:

K T (a) = K 2oo (Q ia) = a 2 + a 4 H ^— a 6 H *2Z a 8 ^ a 10

v"/ v w ' tu v 150 ~ 42000 ~ 5292000 ' 5821200000 9535125600000 '

K-LfV^ = /r 2oo fn r") = -J- T 2 4- T 4 - 101 -r 6 - 10289 r 8 , 7249481 10

V' / V w 5 ; 150 ' "r 1200 5292000 5821200000 ^ 47675628000000

Thus zeros on <9f2 repel as in the case of SU(2)-polynomials considered in ||Han| , |BSZ1|| . The
behavior over long distances is illustrated in Figures in the introduction. Note that
K is invariant (only) under the S 1 action, and thus K 2oo (iax, ia%) = -fT 2oo (0, i(a 2 — oci)).
(However, K 200 ^, r 2 ) ^ K 2oo (0, r 2 - n).)

Remark: The function

detA~(0,,a) = l-f^ N2

which appears in the denominator of our formula for the tangential limit correlation, happens
to be the limit pair correlation function for the eigenvalues of random Hermitian matrices
(see |D|, (5.74)]). However, we are unaware of a good interpretation for det A°° in our
setting.

18

BERNARD SHIFFMAN AND STEVE ZELDITCH

References

[Bel] S. Bell, The Cauchy Transform, Potential Theory, and Conformal Mapping, Studies in Adv. Math.,

CRC Press, Boca Raton, FL, 1992.
[Bcr] S. Bergman, The kernel function and conformal mapping, second revised edition, Mathematical

Surveys 5, American Mathematical Society, Providence, RI, 1970.
[BD] P. Bleher and X. Di, Correlations between zeros of a random polynomial, J. Statist. Phys. 88 (1997),

269-305.

[BR] P. Bleher and R. Ridzal, SU(1, 1) random polynomials, J. Statist. Phys. 106 (2002), 147-171.

[BSZ1] P. Bleher, B. Shiffman and S. Zelditch, Universality and scaling of correlations between zeros on
complex manifolds, Invent. Math. 142 (2000), 351-395.

[BSZ2] P. Bleher, B. Shiffman, and S. Zelditch, Correlations between zeros and supersymmetry, Comm.
Math. Phys. 224 (2001), 255-269.

[Ca] T. Carleman, Uber die Approximation analytischer Funktionen durch lineare Aggregate von
vorgegebenen Potenzen, Ark. Mat. Astr. Fys. 17 (1922-23).

[De] P. Deift, Orthogonal polynomials and random matrices: a Riemann-Hilbert approach. Courant Lec-
ture Notes in Mathematics, 3. New York University, Courant Institute of Mathematical Sciences,
New York, 1999.

[DPSZ] A. Dembo, B. Poonen, Q.-M. Shao and O. Zeitouni, Random polynomials having few or no real
zeros, e-print archive, math. PR/00061 13] .

[EK] A. Edelman and E. Kostlan, How many zeros of a random polynomial are real? Bull. Amer. Math.
Soc. 32 (1995), 1-37.

[EO] P. Erdos and A. C. Offord, On the number of real roots of a random algebraic equation, Proc.

London Math. Soc. 6 (1956), 139-160.
[ET] P. Erdos and P. Turan, On the distribution of roots of polynomials, Ann. of Math. 51 (1950),

105-119.

[Ham] J. M. Hammersley, The zeros of a random polynomial, Proceedings of the Third Berkeley Symposium
on Mathematical Statistics and Probability, 1954-1955, vol. II, pp. 89-111, University of California
Press, Berkeley and Los Angeles, 1956.

[Han] J. H. Hannay, Chaotic analytic zero points: exact statistics for those of a random spin state, J.
Phys. A 29 (1996), L101-L105.

[IM] I. A. Ibragimov and N. B. Maslova, The mean number of real zeros of random polynomials. I.
Coefficients with zero mean, (Russian, English summary) Teor. Verojatnost. i Primenen 16 (1971),
229-248; English translation, Theor. Probability Appl. 16 (1971), 228-248.

[Kl] M. Kac, On the average number of real roots of a random algebraic equation, Bull. Amer. Math.
Soc. 49 (1943), 314-320.

[K2] M. Kac, On the average number of real roots of a random algebraic equation, II, Proc. London Math.
Soc. 50 (1949), 390-408.

[LO] J. E. Littlewood and A. C. Offord, On the number of real roots of a random algebraic equation, I,
J. London Math. Soc. 13 (1938), 288-295; II, Proc. Cambridge Philos. Soc. 35 (1939), 133-148; III,
Rec. Math. [Mat. Sbornik] 12(54), (1943), 277-286.

[SV] L. Shepp and R. Vanderbei, The complex zeros of random polynomials, Trans. Amer. Math. Soc.
347 (1995), 4365-4384.

[ShZl] B. Shiffman and S. Zelditch, Distribution of zeros of random and quantum chaotic sections of positive

line bundles, Comm. Math. Phys. 200 (1999), 661-683.
[ShZ2] B. Shiffman and S. Zelditch, Random polynomials with prescribed Newton polytope, I, e-print
archive,

math. AG/0203074

[SL] V. I. Smirnov and N. A. Lebedev, Functions of a complex variable: Constructive theory, M.I.T.

Press, Cambridge, MA, 1968.
[So] M. Sodin, Zeros of Gaussian analytic functions, Math. Res. Lett. 7 (2000), 371-381.
[Su] P. K. Suetin, Polynomials orthogonal over a region and Bieberbach polynomials. Translated from

the Russian by R. P. Boas. Proceedings of the Steklov Institute of Mathematics, No. 100 (1971).

American Mathematical Society, Providence, R.I., 1974.

EQUILIBRIUM DISTRIBUTION OF ZEROS OF RANDOM POLYNOMIALS

19

[Szl] G. Szego, Uber orthogonalc Polynomc, die zu einer gegebenen Kurve der komplexen Ebene gehoren,
Math. Zeit. 9 (1921), 218-270.

[Sz2] G. Szego, Orthogonal polynomials, fourth edition, American Mathematical Society, Colloquium Pub-
lications, Vol. 23, American Mathematical Society, Providence, RI, 1975.

Department of Mathematics, Johns Hopkins University, Baltimore, MD 21218, USA
E-mail address: bshiffman@jliu.edu, szelditch@jhu.edu

```