arXiv:0902.1027v1 [math-ph] 6 Feb 2009

Condensation of the roots of real random polynomials on the real axis 1

Grégory S hehr , Satya N. Majumdar

2

Laboratoire de Physique Théorique, Université de Paris-Sud, 91405 Orsay Fran e Laboratoire de Physique Théorique et Modèles Statistiques, Université Paris-Sud, Bât. 100, 91405 Orsay Cedex, Fran e 1

2

We introdu e a family of real random polynomials of degree n whose oe ients ak are symmetri independent Gaussian variables with varian e α ha2k i = e−k , indexed by a real α ≥ 0. We ompute exa tly the mean number of real roots hNn i for large n. As α is varied, one nds three dierent phases. First, for 0 ≤ α < 1, one nds that hNn i ∼ ( π2 ) log n. For 1 < α < 2, there is an intermediate phase q where hNn i grows algebrai ally with a ontinuously varying 2 α/2 exponent, hNn i ∼ π α−1 . And nally for α > 2, one nds a third phase where α n hNn i ∼ n. This family of real random polynomials thus exhibits a ondensation of their roots on the real line in the sense that, for large n, a nite fra tion of their roots hNn i/n are real. This ondensation o

urs via a lo alization of the real roots around  the values ± exp α2 (k + 21 )α−1 , 1 ≪ k ≤ n.

Abstra t.

2

Condensation of the roots of real random polynomials on the real axis

1. Introdu tion

Sin e the early work of Blo h and Pólya [1℄ in the 30's, the study of random algebrai equations has now a long story [2, 3℄.

In the last few years, it attra ted a renewed

interest in the ontext of probability and number theory [4℄, as well as in the eld of quantum haos [5℄.

Re ently, we showed that there are also interesting onne tions

between random polynomials and persisten e properties of physi al systems [6, 7℄. Here we onsider real random polynomials, i.e.

oe ients, of degree

n.

polynomials with real random

While these polynomials have exa tly

plane, the number of roots on the real line

Nn

n

roots in the omplex

is a random variable. One would like to

hara terize the statisti s of this random variable and a natural question is thus : what is the mean number

hNn i of real roots and how does it behave

question has been widely studied in the past for Ka 's where

ak

n for large n [4℄? This Pn k polynomials Kn (x) = k=0 ak x with

are independent and identi ally distributed (i.i.d.) random variables of nite 2 2 2 varian e hak i = σ . In that ase it is well known that hNn i ∼ log n, independently of σ . π

This result was rst obtained by Ka [8℄ for Gaussian random variables and it was later shown to hold also for a wider lass of distributions of the oe ients

ak [2, 3℄.

Interesting

generalizations of Ka 's polynomials have been studied in the literature where ak are 2 d−2 , where d > 0 is independent Gaussian variables but non identi al, su h that hak i = k p −1 a real number, leading to hNn i ∼ π (1+ d/2) log n [7, 9℄. Given the robustness of this

hNn i, it is natural to sear h for random polynomials whi h hNn i in reases faster than log n, for instan e algebrai ally. One su h instan e is provided by the real Weyl polynomials Wn (x) dened by n X xk ǫk √ , Wn (x) = (1) k! k=0

asymptoti logarithmi behavior of

for

ǫk√are i.i.d. random variables of zero mean and unit varian e. Thus here, ak = ǫk / k! and the varian e is ha2k i = 1/k!, whi h for large k behaves as ha2k i ∝ e−k log k . 1/2 For these real polynomials in Eq. (1), it is known that hNn i ∝ n . For instan e, in the spe ial ase where ǫk are Gaussian random variables of unit varian e, one has √ hNn i ∼ π2 n [7, 10℄. Another interesting and intriguing instan e of real random where

polynomials was introdu ed a long time ago by Littlewood and Oord [11℄ who studied

Ln (x) given by n xk 1 X ǫk , Ln (x) = + 2 k=1 (k!)k

the random polynomials

(2)

where ǫk = ±1 with equal probability. Thus in this ase 2 2k 2 is hak i = 1/(k!) , whi h behaves for large k as hak i methods, they showed that su h polynomials

ak = ǫk /(k!)k 2 ∝ e−2k log k .

and the varian e Using algebrai

Ln (x) have all their roots real and therefore

hNn i = n.

We thus have here two examples of real random polynomials in Eq. (1) and Eq. (2)

where, at varian e with Ka 's polynomials,

hNn i

grows algebrai ally with

n.

In the

se ond example (2), the number of real roots is ma ros opi  in the sense that, for

3

Condensation of the roots of real random polynomials on the real axis

hNni ∼ nα/2

hNni ∼ log n

0

1 Ka

hNni ∼ n

α

2 Weyl

Littlewood-Oord

Asymptoti behavior of the mean number of real roots hNn i of Pn (x) in Eq. (3) as a fun tion of α. These polynomials exhibit a ondensation of their roots on the real axis for α ≥ 2. Figure 1.

n,

large

Ln (x)

there is a nite fra tion

hNn i/n

of the roots whi h are on the real axis. For

in Eq. (2) this fra tion is exa tly one. We thus say that there is a ondensation

of the roots on the real line, similar to a Bose-Einstein ondensation where a nite fra tion of the parti les of a quantum-me hani al system (Bosons) ondense into the lowest energy level. In the ase of random polynomials, the roots play the role of the parti les and the equivalent of the ground state is the real line. The purpose of this paper is to understand what types of polynomials lead to this

ondensation phenomenon.

Of ourse, it is very di ult to address this question for

ak .

However, guided by the two examples above in Eq. (1) 2 and Eq. (2), and in parti ular by the large k behavior of hak i, we introdu e a family of any random oe ients

Pn (x) indexed by a real α ≥ 0 n X α Pn (x) = ak xk , ha2k i = e−k ,

random polynomials

dened by

(3)

k=0

are real independent Gaussian random variables of zero mean. While α = 0 2 −k log k

orresponds to Ka 's polynomials, we re all that, for Wn (x) in Eq. (1), hak i ∝ e 2 −2k 2 log k and for Ln (x) in Eq. (2), hak i ∝ e . Therefore, due to the extra logarithmi where

ak

fa tor, these random polynomials are not exa tly of the form introdu ed above (3). + However, for α → 1 , one expe ts to re over the behavior of Wn (x) in Eq. (1) while for α → 2+ , one expe ts Pn (x) to behave similarly to Ln (x) in Eq. (2) : this is depi ted s hemati ally in Fig. 1.

α ≥ 0 is varied one nds phase orresponds to 0 ≤ α < 1, where one nds se ond one, orresponding to 1 < α < 2, one has

Our main results an be summarized as follows. three dierent phases. that

The rst

hNn i q ∼ (2/π) log n.

In the

As

2 α−1 α/2 n . And in the third phase, for α > 2, one nds hNn i ∼ n. The π α

ondensation of the roots on the real axis thus happens for α ≥ 2 and as one in reases

hNn i ∼ α,

the ondensation transition sets in at the riti al value

αc = 2.

Furthermore, one

nds that these real roots ondense into a quasi-periodi stru ture su h that there is, α α−1 m on average, one root in the interval [−xm+1 , −xm ] ∪ [xm , xm+1 ], with xm = e 2 , with

1 ≪ m < n.

These dierent behaviors are summarized in Fig. 1.

with phase transitions of statisti al systems the ase a high-temperature phase whereas

α>2

0 1,

This hange of variable (9) is motivated by

e−k xk y k

k=0

while for

essentially governed by the behavior of

α < 1, φ(u, x) has u = u∗ (x) given by

parti ular, for minimum for

φ(u, x)

u (and xed x). In α > 1, it has a single

as a fun tion of

a single maximum while for

∂u φ(u∗ (x), x) = 0 , ∂u2 φ(u∗(x), x) = α(α − 1)u∗ (x)α−2 > 0 , 1  α−1  2 ∗ . log x u (x) = α

(11)

Y = u∗ (x). As a

onsequen e, the density behaves quite dierently in both ases α < 1 and α > 1. + are lo ated in [1, ∞]. For xed x > 1, For α < 1, most of the real roots on R φ(u, x) as a fun tion of u in the interval [0, n] has a global minimum for u = n. Therefore, the sum entering in the expression of cn (x) in Eq. (10) will be dominated by the terms with k ∼ n. The expansion of φ(k, x) in Taylor series around k = n yields The new variable

Y

introdu ed above in Eq.

(9) is thus pre isely

φ(k, x) = φ(n, x) + (k − n)(αnα−1 − 2 log x) + · · · = (1 − α)nα − k(αnα−1 − 2 log x) + · · · ,

where j

the

higher j

order α−j

∂ φ(n, x)/∂u = O(n

)

terms for

j ≥ 2.

−(1−α)nα

cn (x) ∼ e

an

n X

be

Thus, for

(xe− 2 n α

α−1

α2

In that ase, the behavior of the dis rete sum in Eq.

omputation of (α−2)/(α−1)

(log x)

orresponds to

(18), whi h enters in the ∗ α−2 Indeed, in that ase [u (x)] ∝

ρn (x) (14) is quite dierent. → ∞ for large x and therefore the leading term for large x in Eq. (18) m = 0 if b < 1/2 or m = 1 in b > 1/2. Keeping these leading

ontributions, one has

n X



α(α − 1) 2 ∗ α−2 b u (x) g (k − u (x)) exp (−φ(k, x)) ∝ g(−b) exp − 2 k=0   α(α − 1) 2 ∗ α−2 + g(1 − b) exp − . (1 − b) u (x) 2 ∗

where, again, we have omitted the unimportant prefa tor expansion (26), one obtains

ρn (x) ∼

ρn (x)

(πx) cosh

In terms of the variable

Y,

α(α−1) α−2 Y (1 2

the density

ρˆn (Y = ⌊Y ⌋ + b) ∼

∗ (x),x)

(26)

. Using this large

x

in Eq. (14) as

2 h

e−φ(u



ρˆn (Y )

− 2b)

reads,

i,Y =



1  α−1 2 . log x α

α(α − 1)Y α−2 h i. α−2 (1 − 2b) 2π cosh α(α−1) Y 2

(27)

(28)

9

Condensation of the roots of real random polynomials on the real axis

ρˆn (Y ) for large Y < n given by Eq. (28) : it is ˆn (Y ) exhibits qualitatively very dierent from the ase α < 2 (see Fig. 2 a)). Indeed, ρ 1 for large integers 1 ≪ k < n. The height of these peaks is peaks entered around k + 2 α−2 given by α(α − 1)k /(2π) whereas its width s ales like k 2−α . From ρn (x), one an now ompute the mean number of real roots. As in the ase α < 2 (see Eq. (24) and above), one an show that the main ontribution to hNn i omes α α−1 α α−1 n , −1] and [1, e 2 n ]. One thus has from Eq. (28) from the intervals [−e 2 Z ∞ Z n hNn i = 2 (29) ρn (x) dx ∼ 2 ρˆn (Y ) dY In Fig. 2 ), one shows a sket h of

0

∼

0

n Z X

0

k≫1

and nally

hNn i ∼ n , R∞

where we have used

1

n Z α(α−1) k α−2 X 2 α(α − 1)k α−2 dz h i db ∼ , α(α−1) α−2 π cosh z α−2 (1 − 2b) k − π cosh α(α−1) k 2 k≫1 2 (30)

dz/ cosh z = π . This hNn i ∼ n

−∞ axis, hara terized by the fa t that

ondensation of the roots on the real thus o

urs via the formation of this

quasi-periodi stru ture (see Fig. 2 )). More pre isely, this omputation in Eq. (29) R k+1 ρˆn (Y ) dY ∼ 1 whi h means, going ba k to the original shows that for large k , 2 k

x, that there α α−1 xk = e 2 k .

variable with

6. The spe ial ase

is, on average, one root in the interval

[−xk+1 , −xk ] ∪ [xk , xk+1 ],

α=2

In view of the previous analysis, it is tempting to onsider the fra tion of real roots

Φ = limn→∞ hNn i/n for α > 2. One an

as an order paramater. For

onsidering the ase

α=2

α < 2,

one has

Φ=0

whereas

Φ=1

however interpolate smoothly between these two limiting ases by and introdu ing an additional real parameter

2

ha2k i = e−µk .

µ

su h that (31)

Performing the same algebra as explained in the Appendix, one obtains the same formula ∗ −1 log x. The new variable is thus here Y = µ−1 log x as given in Eq. (14) with u (x) = µ

Y = ⌊Y ⌋+b it is easy to see that the density ρˆn (Y ) is given by for 1 ≪ Y < n  " P∞ #2 1/2 P∞ 2 −µ(m−b)2 −µ(m−b)2 (m − b) e µ m=−∞ (m − b)e  P∞ P∞ ρˆn (Y ) =  m=−∞ − , (32) 2 −µ(m−b) −µ(m−b)2 π m=−∞ e m=−∞ e

and, setting

whi h is thus 1-periodi for all given by Eq. (32). For

µ → 0,

µ.

ρˆ(Y ) for α p= 2 −1 ρˆn (Y ) ∼ π µ/2

In Fig. 2 ), one shows a sket h of

the density is almost onstant and

and the modulation of the density in reases with

For large

µ,

the sum in Eq. (32) is

m = 0 and m = 1 and ρˆn (Y ) α = 2 and repla ing Y α−2 by µ.

dominated by the terms orresponding to a formula similar to Eq. (28) setting

µ.

is thus given by

For the average

10

Condensation of the roots of real random polynomials on the real axis

a)

ρˆn (Y )

ρˆn (Y )

(a.u.)

α=2

(a.u.)

α2

)

k-1 k k+1

Y

: Sket h of ρˆn (Y ) (in arbitrary units) given in Eq. (23) as a fun tion of Y for 1 ≪ Y < n for α < 2. b) : Sket h of ρˆn (Y ) (in arbitrary units) given in Eq. (32) as a fun tion of Y for 1 ≪ Y < n for α = 2. ) : Sket h of ρˆn (Y ) (in arbitrary units) given in Eq. (28) as a fun tion of Y for 1 ≪ Y < n for α > 2. Here k denotes an integer with 1 ≪ k < n. Figure 2. a)

number of real roots one has

√  2µ n , µ ≪ 1 hNn i ∝ π  n, µ≫1,

(33)

whi h shows that this family of real random polynomials (31) interpolate smoothly between the ases

α2

(29).

7. A qualitative argument for the transition at

α=2

This ondensation of the roots on the real axis an be qualitatively understood if one

onsiders the random polynomials (for

x > 0) Pˆn (Y ) = Pn (x)

of the variable

Y,

whi h

one writes as

Pˆn (Y ) =

n X k=0



 1 α α−1 aˆk w(k, Y ) , w(k, Y ) = exp − (k − αkY ) , 2

(34)

a ˆk are i.i.d. Gaussian variables of unit varian e. It is easy to see that the weights w(k, Y ), as a fun tion of k , have a single maximum for k = Y where the se ond derivative α−2 . Thus for α > 2, the weights get more and more peaked around is proportional to k and

11

Condensation of the roots of real random polynomials on the real axis

aˆk is typi ally of order O(1). Therefore, given a ˆn (m) is, for α > 2, dominated by a single term orresponding to large integer m, P k = m. Consequently, the sign of Pˆn (m) is essentially the sign of a ˆm . This in turn implies that, if a ˆm and a ˆm+1 have an opposite sign, Pn (x) has, with a probability lose α α α−1 α−1 m , e 2 (m+1) ]. In the ase where a ˆm and a ˆm+1 have the to 1, a root in the interval [e 2 same sign, the same argument shows that Pn (x) has, with a probability lose to 1, a root α α α−1 (m+1)α−1 , −e 2 (m) ]. One thus re overs qualitatively the result we in the interval [−e 2 had found from the omputation of ρ ˆn (Y ) in Eq. (29) where we have shown that Pn (x) α α α−1 α α α−1 α−1 (m+1)α−1 , −e 2 (m) ] ∪ [e 2 m , e 2 (m+1) ]. has, on average, one root in the interval [−e 2 This shows nally that Pn (x) has, on average, hNn i ∝ nreal roots. this maximum for large

k,

whereas

We also point out that our argument explains in a rather intuitive way the result

obtained by Littlewood and Oord [11℄ for the random polynomials Ln (x) (2). For m these spe i polynomials, dening x0 = 0, xm = m m!, they rigorously proved, using algebrai (and rather umbersome) methods, that if

ǫm ǫm+1 = −1

or in

[−xm+1 , −xm ]

if

ǫm ǫm+1 = 1.

Ln (x)

has a root either on

[xm , xm+1 ]

Our argument gives some insight on

their intriguing result and allows to understand it in a rather simple way.

8. Con lusion

To on lude we have introdu ed a new family of random polynomials (3), indexed by

α. For these random polynomials, we have omputed the mean density of real roots ρn (x) from whi h we omputed the mean number of real roots hNn i for large n. We have shown that, while for 0 ≤ α < 1, hNn i ∼ ( π2 ) log n, the behavior of hNn i for α > 1 deviates signi antly from the logarithmi behavior hara teristi for α/2 whereas for Ka 's polynomials. For 1 < α < 2, we have shown that hNn i ∝ n α > 2, hNn i ∼ n. This family of real random polynomials thus displays an interesting a real

ondensation phenomenon of their roots on the real axis, whi h is a

ompanied by an ordering of the roots in a quasi periodi stru ture : this is depi ted in Fig. 2.

Of ourse, the o

urren e of this transition raises several interesting questions like the behavior of the varian e of the number of real roots for large

n

as

α

is varied. It

would be also interesting to ompute the two-point orrelation fun tion of the real roots, whi h is a rather natural tool to hara terize this periodi stru ture we have found. In view of this, we hope that this interesting phenomenon will stimulate further resear h on random polynomials.

Appendix A. A useful expression for the mean density

ρn (x)

In this appendix, we derive the expression for the mean density (14) starting from Eq. (7). We rst write

cn (x) = e−φ(u

∗ (x),x)

n X k=0

˜

cn (x) = hPn (x)Pn (x)i

e−φ(k,x) ,

ρn (x)

as given in Eq.

as (A.1)

Condensation of the roots of real random polynomials on the real axis

where

u∗ (x)

is the lo ation of the minimum of

∗

u (x) = and



φ(u, x)

12

given in Eq. (11)

1  α−1 2 , log x α

(A.2)

φ(u∗(x), x) = (1 − α)u∗ (x)α (A.3) ∗ α ∗ α−1 ∗ ˜ x) = φ(k, x) − φ(u (x), x) = k − αk[u (x)] φ(k, + (α − 1)[u (x)]α . √ The orrelator Cn (x, y) = cn ( xy) is given by Eq. (A.1) together with Eq. (A.3) where √ √ x is repla ed by xy . All the dependen e of Cn (x, y) in x, y is thus ontained in u∗ ( xy) only. From its denition in Eq. (A.2) one has immediately

√ ∂x u∗ ( xy) =

1 ∗ √ 1 [u ( xy)]2−α , α(α − 1) x

(A.4)

from whi h we obtain a set of useful relations

√ √ 2 ∂x,y φ(u∗ ( xy), xy) = −

1 ∗ √ 1 [u ( xy)]2−α α(α − 1) xy

(A.5)

˜ √xy) = 1 (u∗ (√xy) − k) ∂x φ(k, x 1 √ 1 ∗ √ 2 ˜ ∂x,y φ(k, xy) = [u ( xy)]α−2 . α(α − 1) xy For the omputation of ρn (x) from Eq. (7), it is useful to introdu e the notation, for any fun tion g(k) Pn ˜ √ k=0 g(k) exp (−φ(k, xy)) hg(k)iZ = Pn (A.6) . ˜ √xy)) k=0 exp (−φ(k, √ From Cn (x, y) = cn ( xy) and cn (x) given in Eq.(A.1) one obtains √ √ √ 2 ˜ √xy)∂y φ(k, ˜ √xy)iZ ∂x ∂y log Cn ( xy) = −∂x,y φ(u∗ ( xy), xy) − h∂x φ(k, ˜ √xy)iZ − h∂ 2 φ(k, ˜ √xy)iZ . ˜ √xy)iZ h∂x φ(k, (A.7) − h∂x φ(k, x,y From the above relations in Eq. (A.5), it is readily seen that the rst and the last term in Eq. (A.7) an el ea h other. Using the relation in Eq. (7), one nally obtains the relation given in the text in Eq. (14).

Referen es

[1℄ A. Blo h and G. Pólya, On the roots of ertain algebrai equations, Pro . London Math. So . (3) 33, 102 (1932). [2℄ A. T. Bharu ha-Reid and M. Sambandham, Random Polynomials, A ademi Press, New York, 1986. [3℄ K. Farahmand, in Topi s in random polynomials, Pitman resear h notes in mathemati s series 393, (Longman, Harlow) (1998). [4℄ A. Edelman and E. Kostlan, How many zeros of random polynomials are real ?, Bull. Amer. Math. So . 32, 1 (1995). [5℄ E. Bogomolny, O. Bohigas and P. Leboeuf, Distribution of roots of random polynomials Phys. Rev. Lett. 68, 2726 (1992); Quantum haoti dynami s and random polynomials, J. Stat. Phys. 85, 639. (1996).

Condensation of the roots of real random polynomials on the real axis

13

[6℄ G. S hehr, S.N. Majumdar, Statisti s of the number of zero rossings: from random polynomials to the diusion equation, Phys. Rev. Lett. 99, 060603 (2007). [7℄ G. S hehr, S.N. Majumdar, J. Stat. Phys. 132, 235 (2008). [8℄ M. Ka , On the average number of real roots of a random algebrai equation, Bull. Amer. Math. So . 49, 314 (1943); Erratum: Bull. Amer. Math. So . 49, 938 (1943). [9℄ M. Das, Real zeros of a lass of random algebrai polynomials, J. Indian Math. So . 36, 53 (1972). [10℄ P. Leboeuf, Random Analyti Chaoti Eigenstates, J. Stat. Phys. 95, 651 (1999). [11℄ J.E. Littlewood and A.C. Oord, On the number of real roots of a random algebrai equation. II, Pro . Cambridge Phil. So . 35, 133 (1939). [12℄ A.P. Aldous and Y.V. Fyodorov, Real roots of random polynomials: universality lose to a

umulation points, J. Phys. A: Math. Gen. 37, 1231 (2004)