published: Nucl.Phys.B, 621 [PM], (2002), 643-674 Negative moments of characteristic polynomials of random matrices: Ingham-Siegel integral as an alternative to Hubbard-Stratonovich transformation 2 0 0 Yan V Fyodorov 2 n Department of Mathematical Sciences, Brunel University a Uxbridge, UB8 3PH, United Kingdom J 7 Abstract 1 Wereconsidertheproblemofcalculatingarbitrarynegativeintegermomentsofthe(regu- 4 larized)characteristicpolynomialforN×N randommatricestakenfromtheGaussianUnitary v Ensemble(GUE).Averycompactandconvenientintegralrepresentation isfoundviatheuse 6 of a matrix integral close to that considered by Ingham and Siegel. We find the asymptotic 0 expression for the discussed moments in the limit of large N. The latter is of interest be- 0 causeofaconjectured relation topropertiesoftheRiemannζ−functionzeroes. Ourmethod 6 reveals a striking similarity between the structure of the negative and positive integer mo- 0 mentswhichisusuallyobscuredbytheuseoftheHubbard-Stratonovichtransformation. This 1 shedsanew light on ”bosonic” versus”fermionic” replica trick and hassome implications for 0 the supersymmetry method. We briefly discuss the case of the chiral GUE model from that / h perspective. p - h 1 Introduction t a m Recently there was an outburst of research activity related to investigating the moments and : correlation functions of characteristic polynomials Z (µ) = det µ1 Hˆ for random N N v N N − × Xi matrices H of various types[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13,(cid:16)14, 15, 16,(cid:17)17]. Thereareseveral,notcompletelyindependent,sourcesofmotivationbehindstudyingcharacter- r a istic polynomials. First is the intriguing conjecture relatinglimiting distribution ofthe non-trivial zeroess = 1+it oftheRiemannzetafunctionζ(s),onthescaleoftheirmeanspacing,tothatof k 2 k (unimodular) eigenvalues of large random unitary matrices . This implies that locally-determined statistical properties of ζ(s), high up the critical line Res=1/2, might be modelled by the corre- sponding properties of Z(µ), averaged over the so-called Circular Unitary Ensembles (CUE), i.e. with respectto the normalizedHaar measureofthe groupU(N) ofN N unitary matrices. Such × a line of thought and underlying evidences in favour of the conjecture are explained in detail in the papers by Keating andcollaborators,see [1, 2, 3]. In particular,in[1] the authorsmanagedto evaluate arbitary moments of Z µ=eiθ explicitly: N (cid:12) (cid:0) (cid:1)(cid:12) = Z µ=eiθ(cid:12) 2n = N Γ(cid:12)(j)Γ(j+2n) , lim MN,2n =n−1 j! (1) MN,2n N (Γ(j+n))2 N Nn2 (j+n)! D(cid:12) (cid:0) (cid:1)(cid:12) E jY=1 →∞ jY=0 (cid:12) (cid:12) where Γ(z) is the Euler gamma-function. The moments as above were derived for Re n > 1/2 − but can be analytically continued to the whole complex n plane. The limiting value is presented − 1 for the integer positive n. This should be compared with the conjecture[1, 18] 1 T dt ζ(1/2+it)2n a n−1 j! (log T )n2 (2) n T | | ∼ (j+n)! 2π Z0 j=0 Y for the values of the positive integer moments of the Riemann ζ function as T . Here a is n − → ∞ the number specific for ζ-function[1], but the rest shows universal features common to both the random matrix calculations and ζ function. The parameter 1 log(T ) in the above equation − 2π 2π plays the role of the inverse spacing between the ζ function zeroes at a height T and should be − identified with N/2π of the unitary random matrix calculations[1]. Itisimportanttohaveinmindahighdegreeofuniversalityoftheobtainedresults,asdiscussed intheworkbyBrezinandHikami[4]. Byuniversalityoneusuallymeansinsensitivityofthespectral characteristicsto details of distributions of matrix entries. In particular, the limiting value Eq.(1) of the positive integer moments of the characteristic polynomials for unitary matrices is shared, after an appropriate normalisation, by a broad class of Hermitian random matrices, whose most prominent representative is the Gaussian Unitary Ensemble (GUE). Other quantities like the distribution of the logarithm of the characteristic polynomial, its derivative, etc. enjoyed thorough investigations as well[2, 3, 4, 5]. The results were also extended to the ensemble of unitary symmetric matrices (COE- Circular Orthogonal Ensemble), which are related to statistics of zeroes of the so-called L functions[19]. An updated summary of open − questions on relations between the properties of Riemann zeta function and random matrices can be found at web-page of the American Institute of Mathematics[20]. What concerns negative moments of the characteristic polynomials, an additional interest in calculating them arose because of a conjectured behaviour of the negative moments of the (regu- larized) Riemann zeta function: n2 1 T δ −2n log T dt ζ 1/2+ +it 2π , T (3) T Z0 (cid:12) (cid:18) logT (cid:19)(cid:12) ∼ δ ! →∞ (cid:12) (cid:12) (cid:12) (cid:12) put forward in [34] for 1 (cid:12)δ logT. (cid:12) ≤ ≪ The formula Eq.(1) shows divergency at negative integers n and thus provides one with no explicitanswer. Suchadivergencyisanaturalconsequenceofnecessitytoregularizecharacteristic polynomials by adding a small imaginary part to the spectral parameter µ to avoid singularities due to eigenvalues. When such an imaginary part is comparable with the separation between the neighbouring eigenvalues one again might expect universality of the corresponding expressions. The Section 8 of the work by Brezin and Hikami [4] discusses a possible way of calculating the negative moments of the characteristic polynomials themselves. However, in contrast to the moments of the absolute values those are not divergent and, when taken alone, are insufficient for the sake of comparison with Eq.(3). Originalgoalofthepresentpaperwastoreconsidertheproblemofcalculatingboththenegative integermomentsofthe characteristicpolynomialsandthoseoftheir absolutevalue. We succeeded in the analysis of correlation functions of the (regularized) characteristic polynomials in the limit of large N and obtained: [ZN(µ1)ZN(µ∗2)]−n 2πρ(µ) n2 lim = (4) N→∞h[ZDN(µ1)]−nih[ZN(µ∗2)]E−ni (cid:20)−i(µ1−µ∗2)(cid:21) 2 where the regularizations Imµ > 0,Imµ > 0 as well as the spectral difference ω = Re(µ µ ) 1 2 1 2 − were considered to be of the order of mean eigenvalue spacing ∆ = [Nρ(µ)) 1], with ρ(µ) being µ − the mean eigenvalue density at µ= 1Re(µ +µ ). 2 1 2 Such an expression complements that given in Eq.(1) and is expected to be universal and applicable to the Riemann zeta-function. Indeed, taking into account the nonuniformity of the spectraldensity forGUE the correspondensebetweenthe parametersshouldbe asfollows: ∆ 1 = −µ Nρ(µ) 1 log(T/2π). WeseethattherandommatrixresultEq.(4)andtheconjecturedRiemann ∼ 2π ζ-function behaviour Eq.(3) agree in the overallparametric dependence. Anothermotivationforsuchacalculationcomesfromsomequestionsthataroseinapplications of the randommatrix theoryto chaotic anddisorderedquantumsystems which we shortly discuss below. As is well known, eigenvalues of large random matrices[21, 22] played a prominent role in the development of the field of quantum chaos, see e.g.[23], and in revealing its connections to mesocopic systems [24] as well as to some aspects of Quantum Chromodynamics[25]. Results on moments and correlation functions of the characteristic polynomials of large random matrices related to various aspects of quantum chaotic systems can be found in [13, 7, 8, 14, 9], see also [15, 17] for related studies. Characteristic polynomials for chiral random matrix ensembles are used as a model partition function for the phenomenon of chiral symmetry breaking and as such enjoyed thorough considerations, see [25] and references therein. Intimately connected with the field of quantum chaos is the domain of mesoscopic disordered systems. The paradigmatic example is a single non-relativistic quantum particle moving at zero temperature in a static random potential. The system Hamiltonian is, in essence, equivalent to a matrix with random entries. Moreover, in some limiting case such a matrix belongs to the ”domain of universality” of the classical random matrix theory. This fact is of paramount importance and follows from the seminal Efetov’s work, see the book [26], where the notion of the supermatrix (graded) non-linear σ model was introduced for the first time. The latter tool − alternativetoothertechniquesinthetheoryofrandommatricesexpressesexpectationvaluesofthe (products of) resolventsofrandomoperatorsinterms ofintegralsovergradedmatrices containing both commuting and anticommuting entries. The method proved to be capable of dealing with quantitieslessaccessiblebyothermethodsandturnedouttobeindispensableinestablishinglinks between the theory of random matrices and quantum chaotic/mesoscopicsystems, see[26, 27] and references therein. An alternative technique which enjoyed many applications in theoretical physics of disordered systemsofinteractingparticlesisthe(in)famous”replicatrick”. Supposeonelikestocalculatethe ensembleaverage logZ ofalogarithmofsomequantityZ. Thereplicatrickexploitstherelation: h i logZ =lim 1(Zn 1) and attempts to extract the averaged logarithm from the behaviour of the momentns→0Znn , wi−th n being either positive or negative integer. It is clear that in general the h i limitingproceduresuffersfromnon-uniquenessoftheanalyticalcontinuationand”mathematicians will throw up their hands in horror or despair, while physicists are much intrigued”[29]. Random matrices provide an important testing ground for the replica calculations, with the absolute value Z (µ) ofthe characteristicpolynomialplaying the roleof Z. The advantagehere is that one has N | | a better control on results obtained by the ill-defined recipe comparing them against those known from independent calculations. In particular, the early paper by Verbaarschot and Zirnbauer [28] devoted to the relation between the replica and supermatrix methods revealed inherent problems in the former absent in the latter. They foundthatthe naturalanalyticcontinuationn 0gavetwodifferentanswersfor → the”fermionic”(positivemoments)and”bosonic”(negativemoments”)versionsofthereplicatrick, 3 neither of them coinciding with the known result. In contrast, the latter is correctly reproduced within the supermatrix approach. Very recently the verdict of inadequacy of the fermionic replicas was challenged by Kamenev and Mezard[11] and further elaborated by Yurkevich and Lerner[12]. In particular, Kamenev and Mezarddiscovereda convenientintegralrepresentationfor the integerpositive moments providing one with a better control on analytical structure of the expressions. This allowed them to put forward an ansatz which yielded in the limit n 0 the correct exact (nonperturbative) result → for the GUE matrices, and the correct asymptotic results for other symmetry classes . A critical analysis by Zirnbauer [29] demonstrated in a coherent manner that the proposed ansatz was in no way a well-behaved analytical continuation. Even so, such a critique did not devaluate the recipeitselfbutratherrestricteditsdomainofapplicabilitytoperturbativecalculationsandcalled for further investigations. And indeed, the amended fermionic replica trick immediately found applications in the theory of disordered electronic systems with interactions[31, 32] when it was amongveryfewtoolsactuallyavailable. Letusalsomentionarecentdevelopmentintheframework of the Calogero-Sutherland model inspired by closely related ideas[6]. The discussed new insights in the nature of the fermionic replica left, however, unclear if one couldcomeforwardwithameaningfulamendmentfortheirbosoniccounterpartwithinthecontext of nonlinear σ-model ideas (see, however, [30] for the replica limit in the context of orthogonal polynomials). An additional motivation for the present paper was to try to bridge the gap between the cases of the positive and negativen. Our attempt succeded in discoveringan integralrepresentationfor the negative integer moments which is strikingly close to that obtained by Kamenev and Mezard [11] for the positive ones. Technically, analyticity properties inherent in the negative moments of the absolute value of characteristic polynomials is known to result in the non-compact (”hyperbolic”) nature of the in- tegration manifold for the bosonic nonlinear σ model discovered by Scha¨fer and Wegner[35]. In − standard considerations such a manifold enters via the so-called Hubbard-Stratonovich transfor- mation (see the Appendix D for more details). It came as quite a surprise to the present author that the Hubbard-Stratonovichtransformationturnedout to be not only unnecessary,but played, in fact, a misleading role hiding the simple structure of the negative moments. To reveal that structure one should introduce an alternative route via use of the matrix integral close to one considered by Ingham[36] and Siegel[37] many years ago. Astothereplicalimit,thefactofclosesimilaritybetweenourintegralrepresentationandthose in [11, 12] makes it apparent that very the same KMYL recipe ”works” for the bosonic version in the same way as for its fermionic counterpart. This should not be considered as contradicting the Zirnbauer’s argumentation since both versions of the replica trick are somewhat deficient, in the strict mathematical sense. The result obtained just indicates that accepting one of them we have little reasons for discarding the other. Clearly,ourwayofdealingwithnegativemomentssuggestscertainrevisionofthe supermatrix method whose underlying technical idea is a simultaneous uniform treatment of both types of the moments (positive and negative). In fact, we show that after the disorder average is performed treating ”fermionic” and ”bosonic” sectors differently can be of some advantage. Thestructureofthepaperisasfollows. InthesectionIIweexposeourmethodonthesimplest example of negative integer moments of the characteristic polynomials and analyse the obtained expressions in the limit N . Then in the section III we proceed through the calculation for → ∞ the negative moments of the absolute value of the polynomial (in fact, a correlation function). In the section IV we comment on the replica trick and illustrate our statements by addressingbriefly 4 the case considered in [33] - the chiral GUE model - from that perspective. Finally, in the section Vwepresentthesimplestnontrivialexampleofextentionofourmethodtothe generaltypeofthe correlation(generating)functioncontainingsimultaneouslybothpositiveandnegativemomentsof the characteristic polynomials of GUE/chiral GUE matrices. The open questions are summarized in the Conclusion. Technical details are presented in the appendices. 2 Negative Moments of the Characteristic Polynomial Let Hˆ be N N random Hermitian matrix with characterized by the standard (GUE) joint × probability density: (Hˆ)=CNexp NTrHˆ2, CN =(2π)−N(N2+1)NN2/2 (5) P − 2 with respect to the measure dHˆ = N dH dH dH . Here we use to denote complex i=1 ii i<j ij ∗ij ∗ conjugation and denote: dzdz 2dRezdImz. ∗ ≡ Q Q RegularizingthecharacteristicpolynomialZ (µ)=det µ1 Hˆ byconsideringthespectral N N − parameter µ such that Imµ > 0 one represents negative int(cid:16)eger power(cid:17)s of the determinant as the Gaussian integral: n n n 1 i i [ZN(µ)−n]= (4πi)nN d2Skexp(2µ S†kSk− 2 S†kHˆSk) (6) Z k=1 k=1 k=1 Y X X where for k = 1,2,...,n we introduced complex N dimensional vectors S = (s ,...,s )T k k,1 k,N − so that d2S = N ds ds and T, stand for the transposition and Hermitian conjugation, k i=1 k,i ∗k,i † respectively. Q Denotingby ... theexpectationvaluewithrespecttothedistributionEq.(5)weareinterested h i in calculating the negative integer moments of the two types: KN(1,)n(µ1)= [ZN(µ1)]−n (7) as well as (cid:10) (cid:11) KN(2,)n(µ1,µ2)= [ZN(µ1)ZN(µ∗2)]−n (8) D E assuming the regularization Im(µ ) = Im(µ ) > 0. In particular, when Reµ = Reµ , the latter 1 2 1 2 quantity amounts to the negative moment of the absolute value of the characteristic polynomial. Letusstartourconsiderationwiththesimplestofthe two. Performingthe ensembleaveraging in the standard way one finds for the moments of the first type: n n n 1 i 1 KN(1,)n(µ1)= (4πi)nN d2Skexp2µ1 S†kSk− 8N S†kSl S†lSk  (9) Z kY=1  kX=1 kX,l=1(cid:16) (cid:17)(cid:16) (cid:17) Further introducing a n×n Hermitian matrix Qˆ with the matrix elements Qˆkl =S†kSl the integrand is conveniently rewritten as: i 1 exp µ TrQˆ TrQˆ2 1 2 − 8N (cid:26) (cid:27) 5 The standard trick suggested to deal with the apparent problem of the non-Gaussian integral above is to employ the famous Hubbard-Stratonovich transformation amounting to: 1 N i exp TrQˆ2 = dQ˜exp TrQ˜2 TrQ˜Qˆ (10) −8N − 2 − 2 (cid:26) (cid:27) Z (cid:26) (cid:27) thus trading the integration over n n Hermitian matrices Q˜ for a possibility to perform the Gaussian integration over the vectors×S . Then the resulting matrix integral is amenable to the k saddle-point treatment in the limit N . →∞ However one may notice a possibility of an alternative route. Its starting point is similar to the method employed in [38, 39] where it was suggested to rewrite the integral Eq.(9) introducing the matrix δ distribution as the product of δ-distributions of all relevant matrix elements. Then, − obviously, KN(1,)n ∝ dQˆe−81NTrQˆ2In(Qˆ) (11) Z where n In(Qˆ)= d2Ske2iµ1 nk=1S†kSk δ Qˆk,l−S†kSl (12) Z kY=1 P kY≤l (cid:16) (cid:17) and the δ-distribution for complex variables is understood as the product of the δ-distributions for their real and imaginary parts. From now on we do not take care explicitly of multiplicative constantsinfrontoftheintegrals. Wewillshowhowtorestoretheconstantsonalaterstageusing the normalisation condition. To evaluate the last expression we employ the Fourier integral representation for each of the delta-functions involved and combine the Fourier variables into a single n n Hermitian matrix Fˆ. This allows us to proceed as follows: × n In(Qˆ) ∝ d2Ske2iµ1 nk=1S†kSk dFˆexp(2iTr FˆQˆ − 2i Flk S†kSl ) (13) Z kY=1 P Z (cid:16) (cid:17) Xkl (cid:16) (cid:17) dFˆe2iTr(FˆQˆ) det Fˆ µ11n −N ∝ − Z h (cid:16) (cid:17)i Up to this point our consideration was, in fact, parallel to that employed in [38, 39]. We however suggest to go one step further by noticing that the last matrix integral is quite close to the distinguished one considered originally by Ingham[36] and Siegel[37]: 1 JpI,Sn(Qˆ)=ZFˆ>0dFˆe−Tr(FˆQˆ)hdetFˆip =(2π)n(n2−1)p!(p+1)!...(p+n−1)!detQ−(p+n) (14) where both Fˆ and ReQˆ are positive definite Hermitian of the size n and the formula is valid for p 0. The Ingham-Siegel integral can be viewed as a direct generalisation of the Euler gamma- ≥ function integral: Γ(p+1)q (p+1) = dffpe fq to the Hermitian matrix argument and paved − f>0 − a way to the theory of special functions of matrix arguments which is nowadays an active field of R research in mathematics and statistics, see e.g [40]. 1In fact, Ingham and Siegel considered the setof real symmetricmatrices Fˆ rather than their Hermitiancoun- terparts and found the result: (π)n(n4−1) n Γ p+ k+1 detM−(p+n+21). However, their method is equally k=1 2 applicabletobothcases. Q (cid:0) (cid:1) 6 It is an easy matter to adopt their method to calculating our integral 2 which is a matrix- argument generalisation of the formula: ∞ df(feifµq)N = Γ2(πNi)(iq)N−1eifµ for q > 0 and zero oth- −∞ − erwise, provided Imµ>0. Performing the calculation (Appendix A) we find for N n: R ≥ In,N(Qˆ >0)= dFˆe2iTr(FˆQˆ) det Fˆ µ11n −N =CN,ndetQˆN−ne2iµ1TrQˆ (15) − Z h (cid:16) (cid:17)i n(n+1) with C = in2 (2π) 2 and I (Qˆ) = 0 whenever at least one of the eigenvalues of Q is N,n N n,N Γ(j) N−n+1 negative (we recall our choice Imµ >0). 1 As a result wQe arrive (after rescaling the integration variable: Qˆ 2NQˆ) to the following → integral representationfor the negative integer moments of the characteristic polynomial in terms of the integral over the matrices Qˆ: KN(1,)n =CN(1,)nZQˆ>0dQˆe−N[−iµ1TrQˆ+12TrQˆ2]detQˆN−n (16) provided N n. ≥ The overall constant C(1) can be restored by noticing that for Reµ the moments tend N,n 1 → ∞ asymptoticallytoµ−1nN. Ontheotherhand,itiseasytounderstandthatsuchalimitisequivalent todiscardingthequadraticinQˆ termintheexponentofEq.(16). Theresultingintegralisprecisely the Ingham-Siegel one, Eq.(14), and comparison yields the required constant: CN(1,)n =(−iN)Nn(2π)−n(n2−1) N−11 j! j=N n − As the last step of the procedure we choose eigenvalueQs q ,...,q and the corresponding eigen- 1 n vectorsof(positivedefinite)HermitianmatrixQˆ asnewintegrationvariables. Thiscorrespondsto the change of the volume element as: dQˆ = G ∆2 qˆ n dq dµ(U ) where the factor ∆2 qˆ = n { } i=1 i n { } i<j(qi − qj)2 is the squared Vandermonde determinQant, Gn = (2π)n(n2−1) n1 j! and dµ(Un) j=1 sQtands for the normalized invariant measure on the unitary group U(n). The integrand is obvi- Q ously U(n) invariant and we obtain: n n KN(1,)n(µ1)= det(µ11N −Hˆ) − =C˜(1)N,n dqiqi−n ∆2{qˆ}exp−N A(qi) (17) (cid:28)h i (cid:29) Zqi>0Yi (cid:0) (cid:1) Xi=1 where 1 1 C˜(1) =( iN)Nn and A(q)= q2 iµ q lnq (18) N,n − N−1 j! n j! 2 − 1 − j=N n j=1 − The last integral representatQion is our mQain result for the negative moments of the first type: (1) (µ ), valid for arbitary N > n. One can further play with the formulae for finite N and n, KN,n 1 expressing, for example, the negative moments as n n determinants: × n det(µ 1 Hˆ) − det[Φ ] n (19) 1 N − ∝ jk |j,k=1 (cid:28)h i (cid:29) Φjk = ∞dqqN−nπj(1)(q)πk(2)(q)eN[iµ1q−12q2] , Z0 2Wesuggesttocallsuchanintegral”theIngham-Siegel integralofsecondtype”. 7 where π(1)(q),π(2)(q) are any monic polynomials of degree j in the variable q, compare with the j j case of positive moments in [10]. In practice, however, we are mostly interested in the limit of large matrix sizes where one expects the results to show universality as was discussed in much detail in the Introduction. To extracttheleadingasymptoticsasN whenkeepingmomentordernfixedoneshouldemploy →∞ the saddle-point method and find the saddle points of A(q ). i Before doing this we observe that the structure of the derived expressions show striking simi- larity to those obtained for the positive moments of the characteristic polynomials found in [11], see also [4, 10]: n det(µ11N −Hˆ) n =C˜N(1,)neN2nµ21 ∞ dqi∆2{qˆ}exp−N A(qi) (20) Dh i E Z−∞Yi Xi=1 where 1 1 C˜(1) =( i)NnNn2/2 N,n − (2π)n/2 n j! j=1 and the expression for A(q) is the same as in Eq.(18). Q The only essential difference between the two representations (apart from that in the multi- plicative constants and a slight change of the power of the determinant: N n rather than just − N, which is anyway irrelevant for large N) is the range of integration. For the positive moments one integrates over the whole real axis < q < whereas it is over the positive semiaxis i −∞ ∞ 0<q < for the negative moments. i ∞ Thus, we need to consider the saddle points of A (q). It is convenient for further reference to define µ = µ+ ω +iδ, with µ,ω,δ -real, and consid±er Nω,Nδ to be fixed when N . Then 1 2 → ∞ one can replace µ with µ in the saddle-point calculations. The saddle points are obviously given 1 by equations: 1 q iµ =0 (21) i − − q i where i=1,2,...,n. Each of these equations has two solutions: iµ 4 µ2 q± = ± − (22) 2 p We would like to choose the spectral parameter µ to satisfy µ < 2 in accordance with the idea | | of considering the bulk of the spectrum for GUE matrices of large size. Then only for q+ the real parts are positive and the corresponding saddle points contribute to the integral over the positive semiaxis: q >0. Consequently,among2n possiblesetsofsaddlepoints q1±,...,qn± onlythechoice qˆ+ =diag(q+,...,q+) (cid:0) (cid:1) (23) 1 1 should be considered as relevant. This feature constitutes a considerable difference from the case of positive moments where all 2n saddle-points yield, in principle, non-trivialcontributions, albeit of different order of magnitude in powers of the small parameter N 1. For example, for n = 2K − the leadingordercontributioninthelatercasecomesfromthechoiceofhalfofsaddle-pointstobe 2K q+, the restbeing q , with the combinatorialfactor countingthe number ofsuchsets[4]. − K (cid:18) (cid:19) 8 PresenceoftheVandermondedeterminantsmakestheintegrandvanishatthesaddle-pointsets of the exponent and thus care should be taken when calculating the saddle point contribution to the integral. This part of the procedure uses explicitly the so-called Selberg integral: n n Zn(t)= ∞ dξk (ξk1 −ξk2)2e−2t nk=1ξk2 =(2π)n/2t−n2/2 j! (24) Z−∞kY=1 k1Y<k2 P jY=1 for t > 0, see the paper by Kamenev and Mezard [11] for more details. General points of their analysis are applicable for our case without any modification. Expanding around the relevant saddle-points: q = q+ +ξ and performing the required cal- k k culations we find in a straightforwardway the asymptotic expressions for the negative moments: Nn+n2 KN(1,)n(µ1) = (−i)NnNNn−n22 (Nj2=π−N)1n/n2j!"iµ+p24−µ2# 2 (4−µ2)−n42 (25) − iωNn Q Nn µ2 iµ 4 µ2 exp (iµ+ 4 µ2) 1+ − − × ( 4 − − 2 p2 !) p Theformulafor (1) (µ )whereµ =µ ω iδ canbe obtainedfromthe aboveexpressionby KN,n ∗2 ∗2 − 2 − taking its complex conjugate and changing ω ω. Taking the product of the two expressions → − we finally find: KN(1,)n(µ1)KN(1,)n(µ∗2)=N2Nn−n2 N(−2π1)nj! 2 [2πρ(µ)]−n2exp(cid:26)Nn(cid:20)iπρ(µ)ω−(cid:18)1+ µ22(cid:19)(cid:21)(cid:27) (26) j=N n − h i Q where we used the known expression for the (semicircular) mean density of GUE eigenvalues: ρ(µ)= 1 4 µ2. 2π − This completes the calculation of the denominator in the formula Eq.(4). To find the corre- p sponding numeratorwe proceedto derivation of the analogousexpressionsfor the moments of the second type. 3 Correlation functions for the negative moments of the characteristic polynomials. To this end, we consider the product of the expression Eq.(6) with its complex conjugate at a differentvalueofthespectralparameterandaverageitovertheGUEprobabilitydensity. Fromnow on we use the index σ =1,2 to label the N-componentvectors S stemming fromthe first/second σ set of the integrals. To write the resulting expression in a compact form it is again convenient to introduce 2n×2n Hermitian matrix Qˆ with the matrix elements Qˆkσl1,σ2 = S†σ1,kSσ2,l, with k and l taking the values 1,...,n. In terms of such a matrix we have: n n n i i 1 KN(2,)n(µ1,µ2)∝ d2S1,kd2S2,kexp(2µ1 S†1,kS1,k− 2µ∗2 S†2,kS2,k− 8NTr QˆLˆQˆLˆ ) Z kY=1 Xk=1 Xk=1 (cid:16) (cid:17) (27) 9 where Lˆ =diag(1 , 1 ). n n − Again,thestandardwayistouseavariantoftheHubbard-StratonovichtransformationEq.(10) allowingto convertthe termquadraticinQˆ (quarticinS)to thatlinearinQˆ (quadraticinS)and integrateoutthevectorsS. However,presenceofthematrixLˆ andtherequirementofconvergency of the Gaussian integrals necessitates introducing this time a rather non-trivial domain (the so- called ”hyperbolic manifold”, [35]) for the integration over Q˜, to make such a ”decoupling” well- defined. This problem comprehensively discussed e.g. in [28, 41] makes the whole procedure technically involved. For a good pedagogical introduction see [23], the outline of the procedure is presented in the Appendix D of the present paper. Forthe methodsuggestedinthe presentpapersuchproblemdoesnotariseatall. The 2n 2n matrix Qˆ is a Hermitian positive definite and the whole procedure at this stage does not req×uire any modification. Employing the Ingham-Siegel integral of second type yields in this case: KN(2,)n(µ1,µ2)=CN(2,)nZQˆ>0dQˆe−N[−iTrMˆQˆ+12Tr(QˆLˆQˆLˆ)]detQˆN−2n ,Mˆ =diag(µ11n,−µ∗21n) (28) provided N 2n, with the overall constant ≥ 1 C(2) =(N)2Nn(2π) n(2n 1) N,n − − N−1 j! j=N 2n − Clearly, such a uniform applicability can be consideredQas a technical advantage. Nevertheless hyperbolic structure, in fact, lurks in the expression above and manifests itself at the next stage. Namely, equation Eq.(28) differs from its analogue Eq. (16) in one important aspect: it is now of little utility to introduce eigenvalues/eigenvectors of Qˆ as integration variables. Rather, it is natural to treat Qˆ =QˆLˆ as a new matrix to integrate over. Such (non-Hermitian!) matrices are L just those formingthe mentioned hyperbolic manifold. I find it sensible to discuss their properties explicitly in the Appendix B. They satisfy Qˆ†L = LˆQˆLLˆ, have all eigenvalues real and can be diagonalizedbya(pseudounitary)similaritytransformation: Qˆ =TˆqˆTˆ 1,whereqˆ=diag(qˆ ,qˆ ), L − 1 2 and n n diagonal matrices qˆ ,qˆ satisfy: qˆ > 0, qˆ < 0. Pseudounitary matrices Tˆ satisfy: 1 2 1 2 Tˆ LˆTˆ =×Lˆ and form the group U(n,n) (”hyperbolic symmetry”). † In fact, a more convenient way is rather to block-diagonalize matrices Qˆ as L Pˆ U(n,n) QˆL =Tˆ0 1 Pˆ Tˆ0−1 ,where Tˆ0 ∈ U(n) U(n) (cid:18) 2 (cid:19) × and Pˆ are n n Hermitian, with eigenvalues qˆ , respectively. The integration measure dQˆ 1,2 1,2 L × is given in new variables as [28]: dQˆ =dPˆ dPˆ (q q )2dµ(T) where the last factor L 1 2 k1,k2 1,k1 − 2,k2 is the invariant measure on the manifold of T matrices whose explicit expression is presented for − Q reference purposes in the Appendix C. We therefore arrive to the following expression: (2) dPˆ dPˆ I(Mˆ,Pˆ ,Pˆ ) (29) KN,n ∝ ZPˆ1>0ZPˆ1<0 1 2 1 2 × (q1,k1 −q2,k2)2detPˆ1N−2ndet −Pˆ2 N−2ne−N2Tr(Pˆ12+Pˆ22) kY1,k2 (cid:16) (cid:17) 10