Skip to main content

Full text of "Random Kähler Metrics"

See other formats


RANDOM KAHLER METRICS 

FRANK FERRARI, 1 SEMYON KLEVTSOV 1 - 2 AND STEVE ZELDITCH 3 

1 Service de Physique Theorique et Mathematique, Universite Libre de Bruxelles and 
International Solvay Institutes, Campus de la Plaine, CP 231, 1050 Bruxelles, Belgique 
2 ITEP, B. Cheremushkinskaya 25, Moscow 117218, Russia 
fN| \ ^Department of Mathematics, Northwestern University, Evanston, IL 60208, USA 

frank . f errariOulb . ac . be , semyon . klevtsovOulb . ac . be , zelditch@math . northwestern . edu 
1—5 ; 

p^i Abstract. The purpose of this article is to propose a new method to define and calculate 

path integrals over metrics on a Kahler manifold. The main idea is to use finite dimensional 
spaces of Bergman metrics, as an approximation to the full space of Kahler metrics. Wc 
use the theory of large deviations to decide when a sequence of probability measures on 
the spaces of Bergman metrics tends to a limit measure on the space of all Kahler metrics. 
(j_) '■ Several examples are considered. 

43' 



CV 



> 

iO ' 1. Introduction 

t>' 

—l. . 1.1. The aim of the paper and the basic setup. This article is part of a series (cf. 

-^ [22} [23])) devoted to a new approach to random metrics on a manifold M. It is assumed 

that M is a Kahler manifold (M, J, ujq), of any complex dimension n, with an integral Kahler 
form ujq G H 2 (M,Z) and a compatible complex structure J. Motivation to study random 
Kahler metrics arises in large part from the theory of two dimensional quantum gravity 
[37] . because in this case any riemannian metric is Kahler for some complex structure. More 

/\ ■ generally, it is also of interest in both physics and mathematics as a theory of random surfaces 

and stochastic geometry in any complex dimension. In this article we introduce the general 
probabilistic and geometric foundations. Applications relevant to two dimensional quantum 
gravity are discussed in details in the companion articles 



The novel approach to defining and studying random metrics is based on some relatively 
new ideas in Kahler geometry involved in the Yau-Tian-Donaldson program of relating exis- 
tence of constant scalar curvature Kahler metrics in the class of ojq to stability in the sense of 
geometric invariant theory. The main idea is to approximate the infinite dimensional space 
/C[ wo ] of Kahler metrics in the cohomology class [cu ] of u by finite dimensional spaces Bk of 
"Bergman metrics," which are pullbacks of the Fubini-Study metric on complex projective 
space by Kodaira type holomorphic embeddings defined by using holomorphic sections of 
powers L k of the line bundle L with Chern class C\(L) = [ujq]. Much current work in Kahler 
geometry involves the approximation of the infinite dimensional geometry of /C[ Wo ] by the 
finite dimensional geometry of Bk- The idea of this article is to use this approximation to 
regularize path integrals over JC[ wo i by well-defined integrals over Bk- 

l 



2 FRANK FERRARI, 1 SEMYON KLEVTSOV 1 ' 2 AND STEVE ZELDITCH 3 

The space /Cr^i can be parametrized by the Kahler potentials of Kahler metrics u^ in 
the cohomology class of u>q, 

K [wo] = {0 G C°°(M)/R : ^ := Wo + 090 > 0} (1) 

with the equivalence relation that = 0' if and only if — 0' is a constant. Since the 
riemannian metric associated to the Kahler form is given by ig a i = u a i, we use the same 
term 'metric' both for g and corresponding Kahler form u. We assume that the background 
Kahler form uq is integral, i.e. loq G H l,1 (M, 27rZ). In particular, if there exists a constant 
scalar curvature metric in the class [ojq], it is natural to choose ujq to be that metric. In the 
case of a Riemann surface, one usually parametrizes riemannian metrics in a conformal class 
Conf(po) by a Weyl factor a, with g = e 2a g$ for a background metric g$. The cohomology 
class of ujq is entirely determined by the area, and parametrizing the metrics by the Kahler 
potential instead of a is a simple change of variables given by e 2a = 1 — |A O 0, where A 
is the positive laplacian for the metric g . We refer to [32] for a detailed comparison of the 
relations between Conf(g ) and K.[ wo \. 

Our aim is to define probability measures on JCr^i and to study the geometry and physics 
of the resulting random Kahler manifolds. The natural action functionals in this context 
are those arising in Kahler geometry. For example, the Aubin-Yau and Mabuchi actions, 
which are local in the Kahler potential 0, are singled out, even though in complex dimension 
one they are non-local in the Liouville field a. In [221 [22] we discuss in detail an important 
role these new actions play in the problem of the coupling of non-conformal quantum field 
theories to gravity. 

Our approach relies on some recent developments in Kahler geometry. The space /C[ Wo ] is 
formally an infinite dimensional symmetric space of non-positive curvature, when equipped 
with a riemannian metric introduced by Mabuchi [3T] , Semmes [5H] and Donaldson [IB] . In a 
very strong sense, )C\ U0 \ is the infinite dimensional limit of the finite dimensional symmetric 
spaces Bk of Bergman metrics. To be precise, Bk can be identified with positive-definite 
hermitian matrices of rank Nk and determinant one, i.e. with the coset SL(Nk,C)/SU(Nk). 
Here Nk is the dimension of the space H°(M, L k ) of holomorphic sections of L . The approx- 
imation of the differential geometry of K.[ WQ ] by that of Bk is now a fast developing theory, 
see e.g. [TTl HH1 [331 E31 EH1 ESI HQl EH1 El]- We take this theory in a new direction by using 
integrals over Bk to regularize, or rigorously define, integrals over /C[ Wo ]. At the same time, 
the identification Bk — SL(Nk, C)/SU(Nk) makes an interesting direct link between random 
metrics and random matrix models. Let us emphasize that these matrix models are of a 
novel type and are not directly related to the matrix models used to study two-dimensional 
gravity in [U HJJ [231 [I] • 

1.2. Defining path integrals. In physics, path integrals must be defined by suitably regu- 
larizing and renormalizing formal infinite dimensional measures e~ s ^T>g. The main idea of 
this article is to define and compute such path integrals over metrics by finding a sequence 
of measures //& = e~ Sk ^T>kg on the finite dimensional space Bk which tends to e~ s ^T>g in 
the sense that 

/ F(g)e- s ^Vg := lim / F k (g)e- s ^V k g (2) 



RANDOM KAHLER METRICS 3 

for suitable functionals F on /C^j. The parameter k, or equivalently N k , thus plays the role 
of the UV regulator. 

Although we do not put the emphasis on applications to quantum gravity in the present 
work, let us mention that in this case a fundamental constraint on the measure e~ s ^T>g is 
background independence, i.e. independence with respect to the choice of Uq in our context. 
A remarkable feature of the regularization scheme (J2]) is that the spaces B k possess manifestly 
background independent measures T> k g for finite k, which converge to an infinite-dimensional 
background independent measure T>(j). This is used in the companion papers [22| [23]. 

There are two natural ways to obtain measures on the space /C^] of Kahler metrics. In 
the top-down approach, we start with an action functional S of the space K.[ Uo \, which is 
suggested by a particular problem. For example, we show in [221 [23] that the Mabuchi 
action enters naturally in the quantization of a non-conformal matter/gravity system. We 
then need to determine the approximating (regularized) actions Sk- The simplest approach 
is just to restrict S to the Bergman metrics B k C K.[ uo \- But there may be simpler and 
more natural approximations. For example, in §3.31 we propose an approximation scheme 
based on the so-called contraction principle in the theory of large deviations. Let us note 
however that in general the finite dimensional approximations can be quite complicated and 
the corresponding large k limit can be hard to study. 

Therefore, it is also fruitful to consider a bottom-up approach where we choose sequences of 
measures fi k = e~ Sk ^T> k g which are simple from the point of view of our formulation in terms 
of Bergman metrics, and study their large k limits. It is sometimes possible to determine 
the asymptotics of correlation functions of these sequences, so they provide computable toy 
models for the general framework. 

However, in general, it is very difficult to determine asymptotics of correlation functions of 
the \i k . One can then focus on a simpler problem: does the sequence fi k have an asymptotic 
action, or in mathematical terms, does the sequence \x\. satisfy a 'large deviations principle' 
(LDP)? The rate function of the LDP is the asymptotic action. When we construct the 
sequence //& from the bottom-up approach, we usually do not know in advance what infinite 
dimensional theory it tends to. When the sequence satisfies an LDP with a rate function S, 
we know that it tends in weak sense to a theory with action S. 

The simplest sequences \x\. are those which are invariant under the unitary group SU(N k ) 
which acts on Bk — SL(N k , C)/SU(N k ) on the left. Such measures can be reduced to stan- 
dard integrals over eigenvalues. As explained in §4.1[ the [/(AQ-invariance of such measures 
and the reduction to eigenvalue integrals correspond to an infinite dimensional symmetry 
of JC[ Uo ] and the quotient by the symmetry is the space M(M.) of probability measures on 
JR.. This reduction is a new observation in Kahler geometry and is related to the study of 
geodesies of /C^] [38]. In §|U we discuss two sequences /ik of this type. Note that in all cases, 
S'[/(Afc)-invariance leads to background-dependent limits, typically to measures which are 
concentrated at or around the background metric. 

1.3. A few examples of interesting questions. Once a notion of 'random metric' is 
defined, one would ideally like to sample the ensemble of random metrics and graph the 
typical random manifold. Some natural geometric questions are: what is the 'shape' of 



4 FRANK FERRARI, 1 SEMYON KLEVTSOV 1 ' 2 AND STEVE ZELDITCH 3 

a random manifold? How smooth is it? What is the expected volume of a given subset 
U C Ml What is the expected distance between points or the expected diameter of the 
manifold? 

These questions are often related to the computation of correlation functions of natural 
observables on Bk, such as values of the Kahler potential 4>{z) or its derivatives at several 
points (see §E]). Hence we would like to find non-trivial models for which these correlation 
functions are tractable. 

A related circle of problems surrounds scaling relations of random metrics and the KPZ 
formula J2B], recently proved rigorously in |2I], see also the treatment in [TO]. In the Liouville 
theory the KPZ relation can be derived by looking at the scaling of the path integral with 
the area A of the surface [HI d5] • A similar problem in our context is to understand how the 
Mabuchi path integral 

Z k (A n )= f e^^^Vktj) (3) 

JB k (A) 

behaves in the parameters k, A and 7, where A is the area of the surface and Bk{A) is the 
space of Bergman metrics of area A. Convergence of this integral is related to the "stability" 
of (M, L, [u>o], J), i.e. to the growth of the Mabuchi action Sm(uo, 0) along geodesies of )C[ Uo \. 

We plan to study these questions with the formalism introduced in the present paper in 
the future. 

1.4. Plan of the paper. The paper is organized as follows. In §2] we recall the geometric 
background. In £J3], we explain our approach in detail. In subsequent sections, we consider 
different types of examples where the methods apply. We define probability measures on 
/C[ Wo ] using sequences {[ik] of probability measures on Bk- We construct many interesting 
sequences \ik which 'converge' to an infinite dimensional limit in the sense that they satisfy a 
large deviations principle (LDP). Starting from a desired action functional S, we explain how 
to construct sequences \ik satisfying an LDP with rate function S, by applying the contraction 
principle. In §4] we study examples of sequences \x\. arising naturally form the identification 
of Bk as a coset space. In the final section we build finite-dimensional approximations to 
geometric functionals using the contraction principle and pointwise convergence formulas. 

2. Geometric background 

2.1. On the geometry of the space K\ Uo u There is a natural riemannian metric on the 
space /Cr Wo i due to Donaldson, Semmes and Mabuchi [T6J [391 EI] defined by 



Gt(6<j> 1 ,6<h)= SfaWMzW, SfadfceTtU-C^iM). (4) 

Jm 

It is shown in [3TJEH1EES] that G^ is of constant sectional curvature, i.e. that (JC^^g^) is a 
symmetric space. It is formally given by JC[ Uo ] = G<c/G where G is the group of Hamiltonian 
symplectic diffeomorphisms of (M, ujq) and where Gc is the complexification of G. The 
associated formal riemannian volume measure D(j) of /Cr^i is thus the infinite dimensional 
Gc-invariant Haar measure. 



RANDOM KAHLER METRICS 5 

The infinite dimensional geometry is quite formal and many features of finite dimensional 
symmetric spaces do not exist for K,\ Wo \. For instance, G<c does not exist as a group [16], 
although it is formally the group Diff ' of diffeomorphisms / of M so that f*coo is of type 
(1, 1) with respect to the fixed complex structure J. Moreover, the exponential map of K\ Uo \ 
is not well-defined [38], and in directions where it is defined it is geodesically incomplete. 
However, the situation is saved to a large extent by the existence of the finite dimensional 
approximations Bk to JC[ U0 ] that we describe in details later in this section. 

Let us note that in the case of surfaces, where it is also natural to use the conformal Weyl 
factor a instead of <fi to parametrize the metrics, one can consider the metric 

G <r (8a 1 ,6<x 2 ) = [ 5a 1 (z)8a 2 (z)e 2 °LO . (5) 

Jm 

This metric and its associated formal integration volume form T>a is induced from the well- 
known manifestly diffeomorphism invariant DeWitt-Ebin metric on the space of all riemann- 
ian metrics [121 02] • In terms of the Kahler potential, (J5]) becomes the Calabi metric 



G ff (<tyi,<ty 2 )= / (A^cM^XA^^H, (6) 

Jm 

where A^ is the laplacian in the Kahler metric u^. Since it requires a Green's function to 
recover 5cj) from A^(<50), the two metrics (Jl]) and fl5]) have quite different geometric properties. 
In particular, when endowed with the metric (jl]) the space /Cr Wo ] is negatively curved [T6"|l3"9] 
and when endowed with the metric fl6]) it is positively curved [H |8] . 

2.2. Random metrics and random matrices. Crucial to our approach is the existence of 
finite dimensional approximations Bk to JC[ Uo \. The properties of these approximations arise 
from geometric Kahler quantization. Under the integrality assumption [oj ] G H 2 (M,2ttZ), 
there exists a holomorphic line bundle L — >• M with Ci(L) = [uj ]. The space /C^j is thus the 
same as the space % of hermitian metrics h on L with curvature (1,1) form ujh = —idd log h G 
/C[ wo ]. and we may identify Kahler metric u^ G /C[ Wo i with the hermitian metric h = e^^h 
or equivalently with the potential <fi. The Hilbert spaces 'quantizing' (M, u) are defined in 
the standard way to be the spaces H°(M,L k ) of holomorphic sections s a , a = 1, ...,N k of 
L k = L (g) • • • <g> L. We set the number of independent sections to be Nj, = dim H°(M, L k ). 

The Kodaira embedding theorem uses a basis of these holomorphic sections to embed M 
into a high dimensional projective space 

z ^ [ Sl (z), ... ,s Nk (z)} : M ^ CF Nk ~\ 

The projective space has a natural Fubini-Study metric on it. Its pull-back to M is called 
a Bergman metric on M, and it depends on the choice of basis of sections we started with. 
Any two bases can be related by a GL(Nk, C) transformation and the space £>& can be 
identified with the space of all possible bases of sections which lead to inequivalent Bergman 
metrics. Since the Fubini-Study metric is invariant under U(Nk) transformations of sections 
and since the overall normalization of the basis vectors is irrelevant since it does not change 
the associated Bergman metric, we have 

B k ~SL(N kt C)/SU(N k ). (7) 



6 FRANK FERRARI, 1 SEMYON KLEVTSOV 1 ' 2 AND STEVE ZELDITCH 3 

The coset on the right side is the space of positive-definite hermitian matrices of determinant 
one. 

This identification may be viewed in a slightly different way. We can identify B k with the 
space l-i k of all hermitian inner products on H°(M, L k ). The inner product associated with 
a given hermitian metric h on L, which induces a hermitian metric h k on L k , is given by 



Wlb k (h) afi := - \ s a sph k u%, 

v Jm 



where V is the volume, or in two dimensions the area V = A, of the manifold M. Having 
fixed a background metric u G /Cr Wo ], we denote the associated hermitian metric on L by 
ho G "H. This h Q thus induces a reference inner product B.ilb k (h ), which we can fix, without 
loss of generality, to be equal to the identity matrix. Any other inner product may then be 
identified with the positive hermitian matrix representing it relative to the background inner 
product. Since for any metric h = h e~^ in % there is corresponding metric u^ G /C[ wo i, we 
thus obtain a map 

Hilb fc (0) : K [uo] -> H k := {Inner products on H°(M, L k )} . (9) 

We can also introduce a map FSj~ which converts an inner product G G 1-L k to a Bergman 
metric in B k G /Cr Wo i as follows. Given an inner product G, one can form a G-orthonormal 
basis {s a } and an associated Kodaira holomorphic embedding. If e denotes a local frame for 
L — y U C M, then any s G H°(M, L k ) may be expressed locally as s = fe® k . The pull back 
of the Fubini-Study hermitian metric on 0(1) — > CP^ -1 is given by 

N k 

FS k : U k -> /C M , F5 fc (G) = -bg^ l/,^)! 2 G i3 fc C /C M . (10) 

Given the background inner product Hilbfc(/io), we let s° : M — > CP^ -1 denote the asso- 
ciated Kodaira embedding. Let A G GL(N k ,C) denote the change of basis matrix from a 
Hilbfc(/i )-orthonormal basis to a G-orthonormal basis s. Then the positive matrix P corre- 
sponding to G is P = A^A and the Kahler potential for the corresponding Bergman metric 
is given by 

Mz) = FS k {P){z) := ±-\og\As°(z)\ 2 . (11) 

This formula is obviously invariant under the U(Nk) action on A on the left. Since also the 
space of potentials is defined modulo constant shifts, we thus have an identification 

FS k : SL(N k , C)/SU(N k ) -»• B k . (12) 

This identification can be viewed as a correspondence between random positive ermitian 
matrices and random Bergman metrics. It allows us to reduce the asymptotic calculation 
of integrals over metrics to matrix integrals. Thus, under the change of variables P G 
SL(N k ,C)/SU(N k ) — )• (f>p G B k , we may rewrite integrals over metrics as integrals over 
matrices 

/ e- s ^F k (<p)V k <f ) = [ e- s ^F k (P)dP. (13) 

JB h JSL(N k ,C)/SU(N k ) 



RANDOM KAHLER METRICS 7 

Different choices of the discretized action Sk(P) = S k (<f>p), the operators F(P) = F k ((j)p) 
and the measure dP = T> k <pp will be considered later in the text. 

Remark: We now have three ways to view the space B k : 

(1) As a subspace B k C /C[ Wo ] of Bergman metrics in the space of Kahler metrics; 

(2) As the space % k of inner products on H°(M, L k ); 

(3) As the coset space SL(N k , C)/SU{N k ). 

The identification of (3) with (2) is just the choice of a background inner product, while 
the identification of (1) and (2) uses the FS k map (fl2|) . 

Any Kahler metric on M can be approximated by a Bergman metric with increasing 
accuracy as k — > oo. This theorem follows from the large k limit of the asymptotics of 
Bergman or Szego kernels (HI El 03]. A key statement is that for any <fi G /C[ Wo ], 

limF^oHilb fc (0) = 0. (14) 

k— >oo 

This limit formula may be differentiated any number of times. There is even a complete 
asymptotic expansion for the left side in terms of curvature invariants of the right side, 
which follows from the following result. If {s a } is an orthonormal basis with respect to the 
inner product (jHJ), then 



N k 
Pk = 



N k 

Y, s a {z)s a {z)h k = k n + -k n ~ l R{uj h ) + 0(k n ~ 2 ), (15) 

where for large k there is a complete asymptotic expansion on the right, depending on various 
curvature invariants [4~3l |6j [29], see also [19]. The left side of (iT5|) is the contraction of the 
diagonal of the Szego projector 

U h k(z,w) = Ys a (z) <g)s a (w)*, (16) 

a 

i.e. the kernel of the orthogonal projection from all square integrable sections of L k — > M 
to the space H°(M, L k ) of holomorphic sections, with respect to the inner product (151) . The 
pointwise asymptotics (fT5j) provides an important tool in the problem of approximating the 
formal integrals over /Cr wo i by the integrals (1131) over B k - However, as well as in random 
matrix theory, it requires more than just pointwise limits of integrands to obtain the k — > oo 
limits of matrix integrals. 

2.3. Invariant metric and the Mabuchi-Semmes-Donaldson metric. Recall some 
facts on the group GL(N, C) and its subgroup SL(N, C). The group action on the hermitian 
matrices is defined as 

A ■ P = A ] PA . 

Any matrix g G GL(N, C) may be expressed uniquely as the product g = kan of a matrix 
k e U(N), a G D(N), the subgroup of diagonal matrices and n G T(N), the group of upper 
triangular matrices with ones on the diagonal. This factorization is known as the Iwasawa 



8 FRANK FERRARI, 1 SEMYON KLEVTSOV 1 ' 2 AND STEVE ZELDITCH 3 

decomposition of GL(N,C). We denote the Lie algebras of D(N) and T(N) by d and t 
respectively. An important point is that the exponential map 

exp 7 : d © t -► GL(N, C)/U(N) (17) 

is a diffeomorphism since GL(N, C)/U(N) has non-positive curvature. Hence, we may define 
the inverse map logP of P G GL(N,C)/U(N). In the case of positive-definite hermitian 
matrices, it is the usual logarithm. 

The GL(N, (C)-invariant metric on the space GL(N, C)/U(N) can be written as tr(P~ 1 5P) 2 . 
With a suitable normalization, the corresponding Haar measure is 

MP)= 7jJpyj dP ll ---dP NN J] dReP jk dlmP jk . (18) 

l<j<k<N 

Here we note that the space of positive-definite hermitian matrices is open in the space of 
all hermitian matrices and has N real diagonal coordinates Pu ■ ■ ■ Pnn an d N(N — 1) off- 
diagonal coordinates RePjk, ImP^, for 1 < j < k < N, so that the real dimension is N 2 . 
The metric ( TlBl) is easily pulled-back to the metric on SL(N,C)/SU(N) by imposing the 
constraint det P = 1. 

Using the map (Q one can show [7] that the invariant metric on £>& = 5'L(A r fc, C)/SU(Nj s ) 
converges to the Mabuchi-Semmes-Donaldson metric (jlj) on /C^j. Indeed, the variation of 
Hilbfc is given by 



<5Hilbfc(</>)a/3 = TT 

V -IM 



-k - l -^)s a sph k e- k4 > 



stu,;. (19) 

Then the invariant metric on Bk can be pulled back to a metric on JC[ UQ ] as follows 

tr(Hilb,T 1 (0)5Hilb fc (0)) 2 = k n+2 J{8<j>f^ + 0(k n+1 ), (20) 



where the expansion on the right follows from the formula (115]) . Thus, the discretized volume 
element T>k<fi in ffT3|) corresponds to the Haar measure ffT8|) . 

2.4. Geometric action functionals. The full measure on /Cr wo ] will always be taken of the 
form e~ s T>(j) for some action S. For example, in the case of surfaces, the Liouville action 

S(g 0l a) = KS L (g Q ,a) + jj e 2a u , S L (g ,a) = / (da A da + aR uj ) , (21) 

Jm Jm 

where Rq is the Ricci scalar and [i the cosmological constant, is a very natural choice. In 
applications to two dimensional quantum gravity, the positive coupling constant K is fixed 
in terms of the central charge of the matter CFT coupled to gravity. The Liouville action 
satisfies the cocycle identities, 

Sl(9o, 9i) = Sl(9i, 9o) , S L (g , g ± ) + S L (g 1: g 2 ) = S L (g , g 2 ) , (22) 

valid for any metrics go, g\ and g 2 in the same conformal class. These identities are funda- 
mental constraints that any effective gravitational action must satisfy. 



RANDOM KAHLER METRICS 



In our framework, it is natural to express the action functionals in terms of the Kahler 
potential, e.g. the Liouville action reads 

^o,0) = / f^log^Aaiog^ + ^iWog^y (23) 

J M \4: uj w 2 ujqJ 

However, and quite remarkably, there are other geometric functionals, satisfying the cocycle 
conditions, that are local in <fi (but not in a in the case of surfaces) that we can naturally 
consider. The most important one is the so-called Mabuchi action [30J (often called "Mabuchi 
energy" in the mathematical literature), which reads 

S M (u , 0) = / -0 J2 ^ A <~ k ~ * E Ric ( w °) A u t A <~ l ~ k + ^ lo § "4 ) > ( 24 ) 

JM \ n k=Q k=Q u J 

where Ric(cuo) is the Ricci (1, l)-form associated with cuo and R the average Ricci scalar given 
by R = h J M Ru2- The equation of motion of the Mabuchi action yields the condition for 
the scalar curvature to be constant, a property shared with the Liouville action restricted 
to a fixed Kahler class in one complex dimension. Another interesting functional, satisfying 
the cocycle identities, is the so-called Aubin-Yau action, 

Say(9o, <t>) = f (~ Y, ^T<90 A 30 A u;- x - k A u k + 0uA . (25) 

Jm \ k=0 n + l / 

In one complex dimension, the above formulas simplify and yields 

S AY (wo, (/>)=[ (-^ + fao) , (26) 

S M (u , 0) = J {^R<pdB<P + <j>(Ruo - Ric(wo)) + ^ log ^\ . (27) 

Notably, in this case the first two terms of the Mabuchi action are proportional to the 
Aubin-Yau action. 



3. Defining path integrals 

3.1. Formal definition. The aim of this section is to explain how to define path integrals 
of the form J e~ s V<fr, where S is a desired action and T><p is the volume form of the Mabuchi- 
Semmes-Donaldson riemannian metric. The infinite dimensional, continuum path integral 
is only a formal expression. We would like to define path integrals over K,\ Wo \ as a limit 
k — > oo of the finite-dimensional, regularized, path integrals over a sequence of measures 
fik = e~ s ^V k <P on B k : 

[ F(0)e- 5W,) £>0 := lim / F fc (0)e~ SfcW P fc . (28) 

To define rigorously the limit on the right side of this equation is clearly a difficult issue, 
related to the renormalization of the theory, a problem that we will only start to address in 
the present work. 



10 FRANK FERRARI, 1 SEMYON KLEVTSOV 1 ' 2 AND STEVE ZELDITCH 3 

Let us comment on the choice of the volume forms on Bk and /Cr Wo i. As we explained in §2J 
the obvious choice is to take Vk<p to be the Haar measure on Bk (TTSjl . i.e. the volume form 
of the symmetric space riemannian metric, which corresponds in the infinite dimensional 
limit to the riemannian volume density associated to the metric @. Another option is to 
use the fact that the exponential map ( TT7I) exp 7 : TiBk — > Bk is a diffeomorphism, and that 
the tangent space is equipped with the natural Euclidean volume form. This volume form 
converges to the tangent space volume form on /Cr Wo i, which is analogous to the flat volume 
form used in the Liouville theory [HJ H5] . Here we will consider both options. 

The main problem is then to construct sequences Sk of finite dimensional actions, so that 
the right side of Eq. (|28|) converges to a finite theory with the desired action S on /C[ wo ]. 
This is the problem of renormalization. Mathematically, we need to explain what it means 
for a sequence of measures \ik on Bk to converge to a measure on /Cr wo ]. Since Bk C JC[ Uo \ we 
may regard the fik as a sequence of measures on /Cr^i which are concentrated on Bk- Hence 
we can ask whether the sequence fik has a weak limit in the sense of probability theory, i.e. 
whether the right side of (128 p converges for some functions F. 

A natural class of correlation functions K,k(zi, . . . , z n ) is given by taking F in Eq. ([25]) to 
be the product of the Kahler potential field <fi at several points, 

„ n 

JC k (z h ...,z n )= / \{<f>{zj))e- s ^V k <f> 

' n 

Jlog s( Zj )Ps( Zj ) ) e~ s ^ p) V k P, (29) 

SL(N k )/SU(N k ) ' 

where in the second line we use ( 111]) and (fT3|) to change variables from the Kahler potential 
to the positive-definite hermitian matrix P and T>kP is an appropriate measure. In some 
simple examples, e~ Sk ^T> k P splits into the Haar measure on U(Nk) and the measure on 
the eigenvalues of P, in which case it is often possible to calculate some of these correlation 
functions at large k. 

3.2. Large deviations approach. A simpler criterion for convergence of a sequence of 
finite dimensional measures fi k = e~ Sk T>k<fi to the infinite dimensional measure e~ s T>(j) is in 
terms of large deviations theory. In fact, a sequence may satisfy an LDP without converging 
in the previous sense, i.e. without all correlators necessarily having finite limits, so the LDP 
may be viewed as a rough check on whether the sequence is defining an infinite dimensional 
theory. 

As has been already mentioned, we may regard the measures fik on Bk as measures on 
JC[u ] which are supported on (i.e. are zero away from) the submanifolds Bk 6 K,[ UQ \. One says 
that the sequence of measures /x& satisfies an LDP with the speed nk and a rate function S 
if, for all balls B(u,S) C JC[ W0 ] of radius 5, one has 

lim — log fj, k (B(u, 6)) = - inf S 1 ^). (30) 

fc-i>oo Uk u><t,eB(u>,8) 

Here, the radius of the ball B(w, 5) is computed with the riemannian metric on K.\ Uo \. Thus, 
a basic criterion for the sequence fj, k to tend to e~ s ^T>(f) is that it satisfies an LDP with the 



RANDOM KAHLER METRICS 11 

rate functional S. This criterion is rather crude since it only determines the action S and 
does not take into account the integration measure T>cf). But this is an advantage when the 
infinite dimensional measure is hard to define and only the corresponding action is known. 

It may be hard to compute the rate function by integrating over balls. However, one could 
use a simpler method, based on Gartner-Ellis theorem, see [TT] for background. Suppose 
that there exists a limit logarithmic moment generating function 

W(J) = lim — logE fc {e Wfc) }, 

where <pk is our random variable, and ( , ) is an appropriate pairing, which is case-dependent. 
Then the LDP rate functional is determined by the Legendre transform of the generating 
function 

S(4>) = sup{(j,4>)-w(j)}. 

j 

This method is a finite-dimensional analog of the method of effective action in the quantum 
field theory, where the generating functional is defined as a Fourier-Laplace transform of the 
full quantum measure. 

3.3. Contraction principle. No general geometric method for constructing simple and 
tractable approximating actions Sk to a given geometric action S is known. The most 
straightforward way is to restrict the functional S to Bk 

S k = S(<l>)\ Bh . (31) 

This is a good approximation in the sense that Sk — S(<f)) as k — > oo. On the other hand, 
the restriction does not simplify the action nor does it arise naturally from the point of view 
of the geometry of Bk- 

In special cases, considered in the last section, such as the Aubin-Yau action or Mabuchi 
action, the pointwise Bergman kernel expansion (TTo^) can be used to construct alternative 
finite dimensional approximations. Other techniques related to the geometric invariant the- 
ory have been used [SJ ESI ESI El] , see also [21]. These approximations are used to construct 
Sk so that Sk(Tk ((/))) — > S(<f>), where 

T k = FSk o Hilb fc : JC [uo] -► Bk • (32) 

It is not clear whether knowing such pointwise approximations is enough to approximate the 
path integrals. 

However, the contraction principle of large deviation theory suggests some new methods 
for constructing finite dimensional approximations. Namely, we may define 

S k (P)= inf 5(0). (33) 

{<j>eic [LJo] ,mib k (<j>)=p} 

It is clear that the Sk of (133]) does tend to S as k — > oo since the fiber of the map Hilbfc(0) 
shrinks to a single point {</>} in this limit. We now explain how ( 133]) is related to the 
contraction principle. 



12 FRANK FERRARI, 1 SEMYON KLEVTSOV 1 ' 2 AND STEVE ZELDITCH 3 

If we are given a formal infinite dimensional measure /i, we can form approximating mea- 
sures by pushing /j forward to B k , meaning 

fji k = (Hilbfc),// , (34) 

where (Hilbfc)* denotes the push-forward under the map (J9j). It is equivalent but often more 
natural geometrically to consider instead the pushforward 

Hk = {Tk)*li>- (35) 

We observe that 

lim /i fc = /i , (36) 

k— »oo 

so that this method of obtaining finite dimensional approximations is devised to automat- 
ically reproduce the original measure in the large k limit. Indeed, the statement ([3"6"]) is 
equivalent to 

lim / f(FS k omb k (<j>)) f jL k = f /(<% 



for suitable functions /. But this holds by (|M|) as long as the limit can be taken under the 
integral sign since 

lim f(FS k o Hilb fc ((^)) = f(<j>) 

k— >cx> 

for all continuous functions /. 

The contraction principle (see e.g. [TTl [26] ) says that given a sequence of measures \i k 
satisfying an LDP with speed n k and rate function /, and a map T : X — > Y , the pushforward 
measures T*fi k also satisfy an LDP with speed n k and the rate function 

J(y)= inf /(*). (37) 

{x£X,T(x)=y} 

Our proposal (133]) does not in fact follow from the contraction principle because we have 
only one measure \x on /Cr Wo i rather than a sequence of measures and also because we have a 
sequence of maps T k between X = /C^i and Y = B k rather than one fixed map T, as in (13T]) . 
Therefore we view (133]) as an approximation scheme suggested by the contraction principle, 
and the formal considerations above only suggest why it is natural. 

This procedure is actually reminiscent of the Wilsonian notion of renormalizing path 
integrals by integrating out higher frequencies, in the sence that if we have fj, k ~ e~ nkI ^T>{cj)) 
then the pushforward measure corresponds to computing the path integral 



JT-HS) 



over all metrics that correspond to the same regularized Bergman metric. 

In principle, we can modify our problem so that the contraction principle does apply. 
For this we need to embed the measure \x and action S into a one parameter family /i e , S £ , 
satisfying an LDP. Then we can apply the map T k to a sequence T k *fi e . We now have two 
parameters k, e. We can then apply the contraction principle with k fixed to the family in e. 

Since there are many ways to embed /i, S into a family /i e satisfying an LDP, we would 
like to choose one which leads to a good approximation as k — > oo. For some special /i, S 



RANDOM KAHLER METRICS 13 

there may be natural parameters one can use for e. We now consider two ways to proceed 
when no such parameters occur. 

First, if the measure /i on JC\ W0 ] is well-defined, we can just replace our path integral 
J e~ s ^T)(j) by J e~^ s ^T>(p. This new family automatically has an LDP with rate function 
S and it directly gives (133]) as the approximate rate function. 

The approximation (133]) can often be difficult to evaluate since it is a constrained varia- 
tional problem. A drastic simplification is to linearize it around the critical metric ip c , in the 
case when the action S(<p) has a unique critical point. With no essential loss of generality 
we may assume S(<p c ) = 0. We then change the action from S to S £ depending on a small 
parameter e by defining S £ (ip) = -S(cp c + y/er]). We expand the right side in powers of e, 

^S(tp c +^ V ) = ^J 77— | Vc r/ + 0(v/£). 

We then change variables (po+^V — > V in the functional integral. We are simply approximating 
the original measure by a gaussian measure on the tangent space T^JC^y We ignore the 
Jacobian factor of powers of e since we also normalize the measure by dividing by its mass, 
which has the same power of e. The resulting integral is then a gaussian integral 



- \ [(S 2 S Vc !,,„}+< J, V )] 



Vr) 



Tth~fc-[LU } 



which satisfies an LDP with rate function S 2 S (fic . Hence we may apply the contraction prin- 
ciple to this sequence (or family) of measures when we push forward by 

DRi\b k (<f) c ) : T^/C^,] -> T mihk{(j)c) B k . 

Here we use that H.i\b k ((j) c + y/er]) = Hilbfc(0) + v / £-DHilb/ c (0)?7-|-O(£:). Since the rate function 
for the family is the quadratic part 62(77) = {S^S^rj, 77), the finite dimensional approximation 
to this rate functional is given by 

S k (5P)= inf 7 2 (77). (38) 

V :DHab k (cj> c )r,=SP 

The result we get is equivalent to the well-known fact that the pushforward of a gaussian 
measure under a linear map is a new gaussian measure. 

This linearizes the problem to the tangent spaces and we would only expect ( |38l) to be a 
good approximation to S near the critical point (p c . 

4. Measures of the matrix model type on Bergman metrics 

In this section we consider the bottom-up approach, namely we take the finite dimensional 
integrals (1131) . that arise naturally from the random matrix side of the identity rather than 
from the geometric side. In particular we consider £/(iVfc)-invariant measures. We emphasize 
that the corresponding actions Sk(<f>) do not necessarily have a good geometric interpreta- 
tion. In this approach geometric features of Kahler metrics are suppressed and only the 
spectral theory of random positive-definite hermitian matrices is retained. Hence one should 
not expect to get a geometric theory of random surfaces arising in the large k limit. Our 



14 FRANK FERRARI, 1 SEMYON KLEVTSOV 1 ' 2 AND STEVE ZELDITCH 3 

purpose is simply to illustrate the LDP's for such measures, and to compute some interesting 
correlation functions. 

4.1. Eigenvalue models. From the point of view of the standard matrix models the sim- 
plest class of measures on Bk has the form 

H(P) = F(A)dP, (39) 

where .P(A) depends only on the eigenvalues A of the matrix P G Bk, and not on the angular 
part U in the decomposition P = UhU* and the measure dP can be decomposed as dAdlf, 
where dU is the Haar measure on U(Nk). However, we should emphasize that even for the 
eigenvalue type measures the computation of geometric correlation functions (129]) will involve 
nontrivial integration over angular variables. 

In principle, the integration in ( TT3"|) goes over the matrices P G Bk which have unit 
determinant. In the actual computations it is easier to use an unrestricted positive hermitian 
matrix P G GL(Nk)/U(Nk) as integration variable. Since the dilatations P — > cP for c > 
do not change the associated Bergman metric, they correspond to a 'gauge' freedom that 
must be mod out. We shall do that in the standard way by inserting the gauge constraint 
det P = 1 in the path integral with a ^-function, as follows, 

/ioo i*ioo 

da [det P] a /i(P) = / dan a (P). (40) 

-ioo J —ioo 

Obviously, the constrained measures are still of the type of Eq. (|39|) . 



4.2. Empirical measures. Recall that, after fixing a background metric Uq, any inner 
product Hilbfc(0) may be represented as a positive hermitian operator Pk((j>) relative to 
Hilbfc(cjo)- In mathematical terms, we can encode Kahler potentials by the empirical measure 
of eigenvalues (that is to say, by the density of eigenvalues) of either |logPfc(0) or P& (</>). 
In the first case we encode the hermitian operator 

X fe (0):=-j-logP,(0) (41) 



by the density 

v x„^){x) = T7- Yl 5 ( x ~ Xfc J') ' ( 42 ) 



.r- 



N k 

N k 

where {xk,j}j=i are the eigenvalues of A^ (</>). This density defines a measure dux^) — 
vx k (<t>)dx G .M(M), where .M(R) is the convex set of probability measures on BL In the 
second case we can do the same for P k . We can then choose our matrix model such that 
at large k the density (j4"2"|) associated with the matrix ^logPfc, or the density associated 
with the matrix P^ itself, has a smooth large k limit. Clearly this would correspond to quite 
different models. But in either case we have an embedding 

u k ojmb k :JC [uo] ->M(R+). (43) 

Under this embedding, we may push forward the given probability measure on K.\ Uo \ to obtain 



a probability measure on M. ( 

K o Hilb fc ).jU = Prob fc G V{M{R)), (44) 



RANDOM KAHLER METRICS 15 

where V(Ai(M.)) is the metric space of probability measures on .M (K+). The rate function 
for the sequence of measures Prob^, if it exists, will then be a functional on A^(M) rather 
than a geometric functional. This approach is common in the theory of large deviations of 
random matrices, see e.g. [26], and comes up naturally in our problem. 

We also remark that the empirical measure map induces a new family of correlation 
functions, the spectral correlation functions on M. n , defined by 

n 

/C fc (/ii, . . . , //„) = E fc JJ f fe (^i) <g> • ■ ■ ® u k (fj, n ). (45) 

i=i 
Here, Vk® v k'~'® v k denotes the product probability measure on M n determined by Uk, i.e. 
its integral against a function /(/xi, . . . , /x n ) is given by J Rn /(/ii, . . . , fi n )^k{f^i) ■ ■ • Vk(Hn)- 

4.3. Wishart ensemble. The Wishart measure is defined by 

/i Q (P)=e^ trP [detP] a+Ar /i^(P). 

Here /ij\r(P) is the Haar measure ffl8l) . Under the decomposition P = UAU* the radial part 
of the Wishart measure is the eigenvalue distribution 

I ^ 

e -PEf =1 A,| A(A) |2jJ AadA . ) (46) 

where A(A) = [j^^Aj — Afc) is the usual Vandermonde determinant. The normalizing 

constant is Z N = J /x(P) = f^° oo dag~ N( - N+a ' 1 Ylj=i^U + l)r(j + a )- We have also mod out 
by the angular integral. 

Now, we would like to perform a large deviation analysis for the sequence of probability 
measures ( j46l) . Since we have U(N) invariance, we only need to study the logarithmic 
asymptotics of integrals over eigenvalues, which is of a type studied in [27] (see Theorem 
2.1). To obtain the large iV limit, it is necessary to embed the spaces M.+ into one fixed 
space. In the context of Wishart models, the density of eigenvalues 

1 N 
dv\(x) := — 2_, $( x ~~ -\j) dx 

i=i 

of the matrix P will have a smooth limit. Positive matrices with the same density of eigen- 
values are conjugate under U(N). Thus we have mapped the symmetric space Vn into the 
space .Mi~(R+) of probability measures on the positive real numbers K + . We then push- 
forward the measure aVjv under this map to define a probability measure P/v on Atf(M+). 
Thus, P/v is supported on the set of measures (l4"6j) in terms of dv\ as 



N 

\A(\)\ 2 e~ 9 ^= lXj TT A a = e ~9 N fu + xdv ^+ aN Iu + 1o s^^x+n 2 t,( Vx ) ,^, 

i=i 
where 

^( u ) = / / log jar — y\dv(x) du(y) 



16 FRANK FERRARI, 1 SEMYON KLEVTSOV 1 ' 2 AND STEVE ZELDITCH 3 

is the standard Coulomb repulsion term between eigenvalues. The coupling constant in 047p 
scales as g = goN, which is the usual 't Hooft's scaling. Then the sequence of measures (1461) 
satisfies a large deviation principle with the rate function 



S(vx) = ~9o / xdu x + E(z/ A ) . (48) 

Jr + 

Also, integration over a imposes the constraint 

logxdu x = 0. (49) 



It is not difficult to solve the saddle point equation associated with (I48p taking into account 
the constraint (|4"9~|) . The result is a smooth eigenvalue distribution v\ that depends on g 
and whose support is an interval of M + containing the point x = 1. Since the details of this 
solution will not be used presently, we shall explain it in a separate publication. 

It is also possible to explicitly calculate the geometric correlation functions (12"9"|) . Here we 
would like to quote the results which will appear elsewhere. We start with the choice of a 
basis of sections s, which is orthonormal with respect to a choice of background metric Hq 
on the line bundle L k , with corresponding background Kahler metric ujq(z) = — IddloghQ. 
The random metric is then parametrized as oj^p){z) = -rdd\ogs(z)Ps(z). The expectation 
value of a single metric in the large k limit can then be easily shown to be equal to the 
background metric 

E k u4z) = i<9<91og \s(z)\ 2 = u (z) + 0(l/k) . 

In fact, it is not hard to show that the same result holds for an arbitrary measure of the 
eigenvalue type ( 1391) 



u HP )(z) n(P) = lu (z) + 0(l/k) 

In this scaling limit the two-point function has a similar behavior 

E k u^(z)u^(y) = u (z)u (y) + 0(l/k) . 

Essentially, these calculations indicate that in the large k limit the Wishart measures con- 
centrate at the Dirac measure 5u Q , i-e. that random metrics concentrate at the background 
metric. 

This result is not contradictory with the fact that the eigenvalue distribution of P has 
a non-trivial smooth limit as indicated above. Indeed, the eigenvalues of Pk(<P) for <p ^ 
grow (or decay) exponentially with k, in such a way that the eigenvalue distribution of X^ 
(J4"T|) - (J4"2|) has a smooth limit. This scaling follows from the large k asymptotics (1201) and the 
fact that the size N k of the matrix grows as k n . 

More precisely, let vx h {^){x) be the density of eigenvalues of X k {<p) and let u^x) be its 
limit when k — > oo. We call dv^ the Szego limit measure of 0. We conjecture that the limit 
exists and is given by 

du^ = (exp^ 1 ^))*^, (50) 

where 



RANDOM KAHLER METRICS 17 

is the exponential map for the Donaldson-Mabuchi-Semmes metric (j3j). Here, 

is the initial tangent vector to the unique geodesic with endpoints 4> , 0. The conjecture 
follows formally from the fact that | logPfc(0) is the (scaled) inverse of the exponential map 
of Hk and that geodesies of %k tend to geodesies of /C[ Wo ] [33J. The authors have verified the 
conjecture in the case of toric varieties. 

These considerations motivate the study of the model of the next subsection. 

4.4. Models on the tangent space. Let us thus use X k defined by 04ip as the basic 
variable. In effect, we are inverting the exponential map ( TT7|) and defining the probability 
measure on the tangent space TiBk of the symmetric space of positive matrices. 

The simplest such measure is the gaussian measure 

fj,(X) = e- 9NtrXtx dX . (51) 

Here dX is the usual Lebesgue measure on the tangent space. The action trX^X is simply 
the square of the geodesic distance in the symmetric space from to exp X. Indeed, the 
exponential map follows the unit speed geodesic from to exp X and its length is ||X||. 
Consequently, the action has the natural infinite dimensional limit 

S((p ,(p) = djc [uol (0o», 

the square of the geodesic distance of /C[ Wo ] from the background metric (or potential) O to 
0. If instead of dX we choose the Haar measure as the integration measure, it is known that 
the distance function ds k on Bk converges to the distance function of /C[ Wo ] [7J [3]. Thus, we 
formally have 

/ F k (<j>)e- d *^*?V k <i>-+ I F(0)e" 4 ["o] (0o ^P0, (52) 

where T>k4> denotes the Haar measure and Dip denotes the Mabuchi measure, as usual. In 
effect, by using r logP^ instead of P&, the limit measure is spread out from a delta function 
at the background metric to a bell curve centered at the background metric. 

We note that the actions do not depend a priori on derivatives of 0, and thus we may 
expect these random potentials to be very singular. Recall that the corresponding metric is 
uq + idd(j). We could smooth out the singularity by adding a term to the action involving 
derivatives of <ft, but for the moment we consider the simple case without derivative terms. 

The LDP rate functional for the model ( 15TJ) was determined in [2] and is given by 

I(u) = g / x 2 dv{x) — Ti(y) + constant . (53) 

Here we would like to discuss the geometric meaning of this functional in the context of 
Kahler geometry. The pullback of ( 1531) to K,[ Uo ] is given by 

J 0) = 9 (expT 1 (f)) 2 d^ -II log | exp" 1 <j>(x) - exp" 1 (f)(y)\d^ (x)d^ (y) . (54) 
Jm J JmxM 

This is new action functional in Kahler geometry. The exponential map is not well-defined 
(see J3H]), but expT 1 is well-defined as the Dirichlet-to-Neumann operator, taking the 



18 FRANK FERRARI, 1 SEMYON KLEVTSOV 1 ' 2 AND STEVE ZELDITCH 3 

solution of the endpoint problem for geodesies to the initial data of the geodesic. It is 
known that expT 1 is Lipschitz continuous, so the integrals are well-defined. If we write 

(j) = expT 1 (f>(x) for the initial velocity, then the functional becomes 



I( ( f ) o^o) = 9 (fydfifa- / log \cj)(x) - (j)(y)\d^ (x)d^ (y). (55) 

JM J JMxM 

The first term is the one obtained in the previous section, and is simply the Hamiltonian of 
the geodesic flow on T*K[ m u which is equal to the energy of the geodesic from O to 0. 

5. Approximation of geometric actions 

5.1. Aubin-Yau action and the contraction principle. In this section, we consider the 
top-down approach, where we construct the random sequences which converge to a given 
geometric functional on K\ Uo \. We have already shown that the invariant metric on Bk 
converges to the Mabuchi metric (J3J), see (120]) . Therefore, getting the approximation to 
actions is the second important ingredient in our construction. 

Here we apply the contraction principle, formulated in §3.31 to the Aubin-Yau action (1261) 
and determine the corresponding sequence of finite-dimensional approximations Sj~, which 
converge to Say in the sense of large deviation theory. 

We claim that the minimizing potential for the variational problem 

S k (P) = inf S AY {4>) 

{06/C [tt , o]) Hilb fc (^)=P6B fc } 

is a Bergman metric 4>(P) G Bk and that 

S k (P) = S AY {<j>{P)) + 0{l/k). 

Thus, to leading order, the finite dimensional approximation to the Aubin-Yau action ob- 
tained in this way is the same as its restriction to Bk- 

We set up the constrained variational problem as a Lagrange multiplier problem. We want 
to find the critical points for 

S Ay (0)+tr(L(Hilb fc (0)-P)), 

where the Lagrange multiplier L is a hermitian matrix. We then get the following equations 
determining the critical point 

1 1 

y[k+ -A J (s a L aP sp)h k e- k ^ = 1 , Hilb fe (0) a/3 = P a(3 . 

Since A^ is a positive definite laplacian on M, the only solution to the first equation is 

{s a L a pSp)h k e- k<t> = V/k. 
Therefore we can solve for in terms of the matrix L, 

V 1 



e -k<p(L) 



k (sLs)hQ 



RANDOM KAHLER METRICS 19 

and plug the solution to the second critical point equation. We get 

1 f s a sp _ 

k J JHFf^ ~ F ■ 

The last two equations imply that <f>(L) = FS k (L), where the matrix L is implicitly deter- 
mined by the matrix P, according to the last equation. It follows that 4>l is a Fubini-Study 
metric, and that to leading order 

S k (P) = S AY (4>(P))+0(l/k). (56) 

Therefore the contraction principle yields the same action as the pull-back of the action 
under the FS k map. 

5.2. Mabuchi action and Donaldson functional. The Mabuchi action ( 12"7|) has a mini- 
mizer <p c corresponding to the constant scalar curvature metric, which always exists in com- 
plex dimension one in each Kahler class. The contraction principle, applied to the Mabuchi 
action, leads to a constrained variational problem that is very hard to solve. 

However, there exists another approximation method, proposed by Donaldson [18], which 
is based on the pointwise Bergman kernel asymptotics ffT5l) . The idea is to consider the 
variation of the log det Hilb with respect to <f), 

51ogdetHilb fc (0) = TrHilbfc V)£Hilb fe (</>) = i / {-kp k (u^) - ^p^u^S^, 

where we have used Eq. ( Tl9|) . Using the expansion Eq. ( flBl . one can then integrate the right 
side and get 

Flog det Hilb fc (0) = -kN k S AY (oooA) + \s M {u Q ,<t>) + 0(1) , 

provided that we chose the basepoint for the Hilb inner product to be the identity matrix, 
as we've always done. Therefore the following functional on JCr Wo ] approximates the Mabuchi 
action on K,\ UQ \ in the large k limit 

£{u , 0) = ^ log det Hilb(0) + 2N k S AY (u , 0) . (57) 

Following Donaldson in [18] , let us now define the following functional on positive hermitian 
matrices P e GL(N k , C)/U(N k ), 

Z{P) = j log det P + N k S AY (w , <t>{P)) ■ (58) 

This functional is obtained by pulling back the first term in ( |57|) from /C[ Wo i to 1-L k using 
the Hilb map and by pushing forward the second term in (137)1 in the opposite direction 
using the FS k map (TT2l . The difference between the two operations is at least of order 1/k 
on B k , as we pointed out in (1141) . Hence, the functional (158 j) defines automatically a good 
approximation to the Mabuchi action on the space B k . 

We can then consider the following sequence of measures on positive-definite hermitian 
matrices in GL(N k ,C)/U(N k ), 

fi k (P)=e- 9Z ^dp N (P), (59) 



20 FRANK FERRARI, 1 SEMYON KLEVTSOV 1 ' 2 AND STEVE ZELDITCH 3 

where g is a coupling constant and we use the Haar measure ([TBI) . Note that, as required, 
Z(P) is invariant under the multiplication of P by a strictly positive constant and so is the 
measure (TT8|) . Therefore (|59|) is well-defined on the space Bk- If we impose explicitly the 
constraint detP = 1, the functional Z(P) becomes simply proportional to the functional 
(|56|) . that we obtained by applying the contraction principle to the Aubin-Yau action. The 
fact that this functional effectively defines the restriction of the Mabuchi functional to the 
space of Bergman metrics has been also observed in [M] . 

The study of the k — > oo limit of (159|) . see also (J3J), is a very interesting open problem 
that we plan to discuss in the future. 

A second approach to regularizing the Mabuchi action is to linearize the problem around 
the constant scalar curvature metrics by applying the contraction principle to the modified 
Mabuchi functional, 

-S M (<p c + y/erj), 

see section [3731 As in the Liouville theory [12], the semiclassical expansion can be developed 
for the Mabuchi action around the critical point and this reduces the Mabuchi action to its 
gaussian approximation 

^S M {<f>c + V^V) = \J Vi&o - 2R)A oV co + ■■■ (60) 

where Ao is the laplacian on the critical metric. The contraction principle leads to the 
following constrained problem, 

S Mtk (5P) = inf If V (A - 2R)A oV u . 

D 4>c mib k (<f>)=sp 8 J 

It would be interesting to see how well this approximates our original problem. 

6. Conclusions 

Given a Kahler manifold M, we have introduced a new way to integrate over the space of 
Kahler metrics in a fixed Kahler class on M. The main ideas of the method are: 

• To use the Kahler potential as the basic dynamical variable. The corresponding 
metric Uq + idd(p is local in <fi. 

• To use Kahler quantization to construct finite dimensional approximations Bk to IC^o] ■ 
The spaces Bk may be identified with symmetric spaces after fixing a background 
metric, hence there are many natural choices of integration measures and of action 
functionals on them. There is a long developed theory of approximating geometric 
functionals on /C[ Wo ] by finite dimensional ones on Bk- 

• To use the theory of large deviations to decide when a sequence \x\. of probability 
measures on Bk tends to a limit measure on /C[ Wo i and to determine the limit rate 
function. 

Several examples were explored in some detail. Our method is very general and we believe 
it could be applied to and shed a new light on a wide range of problems in physics and 
mathematics. 



RANDOM KAHLER METRICS 21 

Acknowledgments. We would like to thank M. Douglas for useful discussions. This 
work is supported in part by the belgian FRFC (grant 2.4655.07), the belgian IISN (grant 
4.4511.06 and 4.4514.08), the IAP Programme (Belgian Science Policy), the RFFI grant 
11-01-00962 and the NSF grant # DMS-0904252. 



[1] 

[2] 

[3] 

[4] 

[5] 
[6] 

[7] 
[8] 

[9 
[10 

[11 

[12 
[13 

[14 
[15 
[16 

[17 
[18 

[19 

[20 



References 

A. A. Belavin and A. B. Zamolodchikov, On correlation numbers in 2d minimal gravity and matrix 



models, J. Phys. A 42 (2009) 304004, |arXiv: 0811 .0450 [he pth]] 



G. Ben Arous and A. Guionnct, Large deviations {or Wigner's law and Voiculescu's non- commutative 
entropy, Probab. Theory Related Fields 108 no. 4 (1997) 517-542. 

Bo Bcrndtsson, Probability measures related to geodesies in the space of Kahler metrics, 
larXiv : 0907 . 1806 [math . DG] [ 

E. Brczin and V. A. Kazakov, Exactly solvable field theories of closed strings, Phys. Lett. B236 (1990) 
144-150. 

S. Calamai, The Calabi's metric for the space of Kahler metrics, arXiv: 1004. 5482 [math. DG] . 

D. Catlin, The Bergman kernel and a theorem of Tian, in Analysis and geometry in several complex 

variables (Katata, 1997), Trends Math., 1-23, Birkhauser Boston MA, (1999) 

X. X. Chen and S. Sun, Space of Kahler metrics (V)- Kahler quantization, arXiv: 0902 .4149 [math.DG] . 

B. Clarke and Y. A. Rubinstein, Ricci flow and the metric completion of the space of Kahler metrics 

larXiv : 1 102 . 3787 [math . DG] | 

F. David, Conformal field theories coupled to 2d gravity in the conformal gauge, Mod. Phys. Lett. 3 no. 
17 (1988) 1651-1656. 

F. David and M. Bauer, Another derivation of the geometrical KPZ relations, J. Stat. Mech. no. 3 (2009) 
P03004, arXiv:0810~28 58 [hep^th]] 

A. Dembo and O. Zeitouni, Large deviations techniques and applications, Applications of Mathematics 
38, Springer- Verlag, New York (1998). 

B. S. DcWitt, Quantum Theory of Gravity. 1. The Canonical Theory, Phys. Rev. 160 1113 (1967). 

D. Ebin, The manifold of riemannian metrics, Proc. Sympos. Pure Math., 15, AMS, Providence RI 

(1970). 

P. Di Francesco, P. H. Ginsparg and J. Zinn- Justin, 2D gravity and random matrices, Phys. Rcpt. 254 



(1995) 1, arXiv:hep-th/9306153[hep-th] 



J. Distler and H. Kawai, Conformal field theory and 2D quantum gravity, Nucl. Phys. B321 no. 2 (1989) 

509-527. 

S. K. Donaldson, Symmetric spaces, Kahler geometry and Hamiltonian dynamics. Northern California 

Symplectic Geometry Seminar, Amer. Math. Soc. Transl. Ser. 2, 196 13-33, AMS, Providence RI, 

(1999). 

S. K. Donaldson, Scalar curvature and projective embeddings. I, J. Diff. Geom. 59 no. 3 (2001) 479-522. 

S. K. Donaldson, Scalar curvature and projective embeddings. II, Q. J. Math. 56 no. 3 (2005) 345-356, 

arXi v : math/0407534 [math . DG] 

M. R. Douglas and S. Klevtsov, Bergman kernel from path integral, Comm. Math. Phys. 293 no. 1 



(2010), 205-230, arXiv : 0808. 2451 [hep-th] 



M. R. Douglas and S. H. Shcnkcr, Strings in less than one dimension, Nucl. Phys. B335 no. 3 (1990) 

635-654. 

[21] B. Duplantier and S. Sheffield, Liouville Quantum Gravity and KPZ, arXiv : 0808. 1560 [math. PR] . 
[22] F. Ferrari, S. Klevtsov and S. Zclditch, Random geometry, quantum gravity and the Kahler potential, 

preprint (2011). 
[23] F. Ferrari, S. Klevtsov and S. Zelditch, to appear in 2011; and works in progress. 
[24] J. Fine, Calabi flow and projective embeddings. With an appendix by Kefeng Liu and Xiaonan Ma, J. 

Diff. Geom. 84 no.3 (2010) 489-523, [arXiv: 0811 .0155 [math. DGJl 
[25] D. Gross and A. Migdal, A nonperturbative treatment of two-dimensional quantum gravity, Nucl. Phys. 

B340 (1990) 333-365. 



22 FRANK FERRARI, 1 SEMYON KLEVTSOV 1 ' 2 AND STEVE ZELDITCH 3 

[26] A. Guionnct, Large deviations and stochastic calculus for large random matrices, Probability Surveys 1 

(2004) 72-172 |arXlv : math/0409277 [mathTPR]] 
[27] K. Johansson, On fluctuations of eigenvalues of random hermitian matrices, Duke Math. J. 91 no. 1 

(1998) 151-204. 
[28] V. G. Knizhnik, A. M. Polyakov and A. B. Zamolodchikov, Fractal structure of 2D-quantum gravity, 

Mod. Phys. Lett. A3 no.8 (1988) 819-826. 
[29] Z. Lu, On the lower order terms of the asymptotic expansion of Zelditch, Amcr. J. Math. 122 no. 2 

(2000) 235-273 [arXiv : math/981 1 126 [math . DG] | 
[30] T. Mabuchi, K-energy maps integrating Futaki invariants, Tohuku Math. J. 38 (1986) 575-593. 
[31] T. Mabuchi, Some symplectic geometry on compact Kahler manifolds, Osaka J. Math. 24 (1987) 227- 

252. 
[32] W. Mullcr and K. Wendland, Extremal Kahler metrics and Ray- Singer analytic torsion, Geometric 

aspects of partial differential equations (Roskildc, 1998), Contcmp. Math. 242 135-160, AMS Providence 

RI, (1999). 
[33] D. H. Phong and J. Sturm, The Monge- Ampere operator and geodesies in the space of Kahler potentials, 

Invent. Math. 166 (2006) 125-149 [arXiv : math/0504157 [math . DG] | 
[34] D. H. Phong and J. Sturm, Lectures on stability and constant scalar curvature, Current developments 

in mathematics, (2007) 101-176, Int. Press, Somcrvillc MA (2009) larXlv: 0801 .4179 [math. DG"H 
[35] D. H. Phong and J. Sturm, Stability, energy functionals and Kahler- Einstein metrics, Comm. Anal. 

Geom. 11 (3) (2003 ) 565-597 |arXiv : math/0203254 [math . DG] | 
[36] D. H. Phong and J. Sturm, Scalar curvature, moment maps and the Deligne pairing, Amer. J. Math. 

126 (3) (2004) 693-712 [arXiv : math/0209098 [math . DG] | 
[37] A. M. Polyakov, Quantum geometry of bosonic strings, Phys. Lett. B103 (1981) 207-210. 
[38] Y. A. Rubinstein and S. Zelditch, The Cauchy problem for the homogeneous Monge- Ampere equation, I. 

Toeplitz quantization, arXiv: 1008 .3577 [math. DG] ; The Cauchy problem for the homogeneous Monge- 

Ampere equation, II. Legendre transform, arXiv: 1010. 2442 [math. DG] 
[39] S. Semmcs, Complex Monge-Ampere and symplectic manifolds, Amer. J. Math. 114 no. 3 (1992) 495- 

550. 
[40] J. Song and S. Zelditch, Test configurations, large deviations and geodesic rays on toric varieties, 

larXiv : 0712 . 3599 [math. DG]] 
[41] G. Tian, On a set of polarized Kahler metrics on algebraic manifolds, J. Diff. Geom. 32 (1) (1990) 

99-130. 
[42] A. B. Zamolodchikov, On the entropy of random surfaces, Phys. Lett. B 117 (1982) 87. 
[43] S. Zelditch, Szego kernels and a theorem of Tian, IMRN 6 (1998) 317-331, [arXiv : math-ph/0002009] 
[44] S. Zhang, Heights and reductions of semi- stable varieties, Compositio Math. 104 (1) (1996) 77-105.