SPhT/92-001

Combinatorics of the Modular Group II

The Kontsevich integrals
arXiv:hep-th/9201001v1 31 Dec 1991

C. Itzykson and J.-B. Zuber

Service de Physique Théorique de Saclay*, F-91191 Gif-sur-Yvette cedex, France

Abstract We study algebraic aspects of Kontsevich integrals as generating functions

for intersection theory over moduli space and review the derivation of Virasoro and KdV
constraints.

Contents

0. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1. Intersection numbers. . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2. The Kontsevich integral. . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1. The main theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2. Expansion of Z on characters and Schur functions . . . . . . . . . . . . 7
2.3. Proof of the first part of the Theorem. . . . . . . . . . . . . . . . . . 13
3. From Grassmannians to KdV . . . . . . . . . . . . . . . . . . . . . . . 19
4. Matrix Airy equation and Virasoro highest weight conditions. . . . . . . . . . 29
5. Genus expansion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
6. Singular behaviour and Painlevé equation. . . . . . . . . . . . . . . . . . 38
7. Generalization to higher degree potentials . . . . . . . . . . . . . . . . . . 41

1/1992 Submitted to Int. J. Mod. Phys.

* Laboratoire de la Direction des Sciences de la Matière du Commissariat à l’Energie Atomique

0. Introduction

The study of two–dimensional gravity has uncovered a rich mathematical structure includ-
ing Virasoro constraints, KdV flows, N = 2 twisted supersymmetry, etc. The remarkable
contributions of Witten [1] and Kontsevich [2] to its topological interpretation have re-
duced an intersection problem on the moduli space of curves to the computation of a
matrix integral over N × N hermitian matrices
R  2

M3
dM exp −tr ΛM
2
− i 6
Z(Λ) = R 2 (0.1)
dM exp −tr ΛM2

While many aspects of this connection have been already discussed by these authors[3][4],
our endeavour has been to study this integral in a purely algebraic context, combining
reviews of former work and new results. We apologize to the expert reader who may skip
sec. 1 where we sketch Kontsevich’s construction and parts of sec. 3 which present a short
account of Sato’s work on τ –functions. In sec. 2 we study the integral (0.1) as an expansion
in powers of the traces trΛ−r . After taking a suitable large N limit, we prove the crucial
property that the asymptotic expansion of this integral does not depend on the even traces
trΛ−2r . This result which followed in Kontsevich’s work from topological considerations is
derived here in a purely combinatorial fashion. The arguments although straightforward
are unfortunately rather intricate since they cannot directly apply to the finite N integral,
for which only N of the traces are algebraically independent. The integral is thus subject
to several constraints:
(i) The equations of the KdV hierarchy pertaining to a differential operator of second
order follow from Sato’s work and this independence with respect to the even traces.
These provide evidence of the equivalence of Kontsevich’s model with the one–matrix
model considered by Witten in [1].
(ii) The Virasoro highest weight constraints follow from the matrix Airy’s equation
satisfied by the numerator of (0.1) (sec. 4) [3].
As for the standard matrix models, a systematic genus expansion is possible, the
leading term of which had been obtained by Witten [1] and from another point of view by
Makeenko and Semenoff [5] relying on earlier work by Kazakov and Kostov [6] (sec. 5). An
analysis of the resulting expressions in a certain singular limit (sec. 6) provides effective
ways to resum families of intersection numbers and derive explicit fomulae. This step
introduces a Painlevé equation and its perturbations.

1
 The integral (0.1) admits a generalization in which the cubic potential is replaced by
a suitable higher degree polynomial [7][2]. The corresponding topological interpretation
presented in a recent paper [8] involves an intersection theory on a finite covering of
moduli space. On the other hand, it is most likely equivalent to the multimatrix integrals.
Our previous discussion extends to these cases without any difficulty of principle (sec. 7)
although the calculations soon become very cumbersome.

1. Intersection numbers.

Witten conjectured in [1] that the logarithm of the partition function of the general one-
matrix model [9][10], expressed in terms of suitable deformation parameters ti could be
expanded as
X tk00 tki i
ln Z = hτ0k0 . . . τiki . . .i ... ... (1.1)
k0 ! ki !
k0 ,...,ki ,...

where the bracketed rational coefficients admitted the following interpretation as intersec-
tion numbers. Let Mg,n be the moduli space of complex dimension 3g − 3 + n ≥ 0 of
algebraic curves of genus g with n marked points x1 , . . . , xn and Mg,n a suitable compact-
ification. The cotangent spaces at xi define line bundles Li with first Chern class c1 (Li ) in-
P
terpreted as 2–forms over Mg,n . For integral non-negative df ’s such that 3g−3+n = f df
the integral Z
c1 (L1 )d1 . . . c1 (Ln )dn (1.2)
Mg,n

(independent of the ordering since 2–forms commute and powers are exterior powers) is
a rational positive number when one considers Mg,n as an orbifold, i.e. the quotient of a
contractible ball (Teichmüller space) by the mapping class group. If ki = #{df = i, i ≥ 0}
P
(so that i≥0 i ki = 3g − 3 + n) then Witten’s conjecture was that hτ0k0 . . . τiki . . .i defined
by (1.1) is given by the integral (1.2). (We have skipped a number of essential technicalities
which make the above definitions sensible). An intuitive picture of the line bundles Li over
Mg,n is not straightforward but at least one can see that hτ03 i = 1, since M0,3 is a point!
1
Even to find that hτ1 i = 24
from the definition is non trivial.
These intersection numbers being topological invariants, Kontsevich has been able to
reduce them to more manageable expressions using a cell decomposition of Mg,n inherited
from the physicists’ “fat–graph” expansion of Hermitian matrix integrals. One considers
connected fat (i.e. double-line) graphs with vertices of valency three or more, genus g and

2
n faces (dual to the n punctures). One assigns to each (double) edge e a positive length ℓe
P
and to each face f a perimeter pf = e⊂f ℓe , where e ⊂ f denotes the incidence relations.
The set of such fat graphs with assigned ℓ’s is a decorated combinatorial model for Mg,n .
Cells have “dimensions” over the reals obtained by counting the number of independent
lengths for fixed values of the perimeters, namely E − n where E is the number of edges.
If Vp denotes the number of p–valent vertices, one has

X
E−n− Vp = 2g − 2
p≥3
X X
2E = pVp ≥ 3 Vp . (1.3)
p≥3

Thus
E − n ≤ 2(3g − 3 + n) (1.4)

with equality (top dimension) if and only if all vertices are trivalent.
Let k be the number of edges bordering a face f and ℓ1 , ℓ2 , . . . , ℓk be their successive
lengths as we circle around the boundary counterclockwise (the faces come with a positive
orientation), up to cyclic permutation. One introduces the 2–form

X    
ℓa ℓb
ω= d ∧d (1.5)
pf pf
1≤a<b≤k−1

invariant under a rescaling and a cyclic permutation of the ℓ’s, which is the first Chern
class of a universal S1 bundle with k–gons as fibers over a combinatorial model for Mg,n .
For fixed values of the perimeters pf the formula obtained for the intersection numbers is
then
Z Y
d
hτd1 . . . τdn i = ωf f (1.6)
f

the integral being on top-dimensional cells (we omit a discussion of the compatibility of
orientations with the complex structure of Mg,n ). A generating function is obtained by
computing
Z ∞ Y  Z (P p2 ωf )3g−3+n
−λf pf f
dpf e (1.7)
0 (3g − 3 + n)!
f

3
where the integral sign stands both for the integral over a cell and a sum over cells of
dimension 3g − 3 + n weighted by the inverse of the order of their automorphism group
(orbifold integration). On the one hand this is

X Z Y
n 2df
−λf pf
pf
hτd1 . . . τdn i dpf e (1.8)
1
df !
d1 +...+dn =3g−3+n
X n
Y
3g−3+n −(2d +1)
=2 hτd1 . . . τdn i (2df − 1)!! λf f .
d1 +...+dn =3g−3+n 1

On the other hand one can proceed to a direct evaluation using

Y X X 3g−3+n
1
dpf ∧ dℓa ∧dℓb = 25g−5+2n dℓ1 ∧. . .∧dℓE (1.9)
(3g − 3 + n)!
f f 1≤a<b≤kf −1
ea ,eb ⊂f

up to an overall orientation which is henceforth ignored. The computation of the above

Jacobian (which depends on the structure of the graph only through g and n) is a delicate
matter for which we refer to [2]. Inserting this expression into the integral one notes
that each edge of length ℓ is shared by two faces f, f ′ ⊃ e, the corresponding integral
2
contributing a factor λf +λ′f
, while the ratio of powers of 2 reads 25g−5+2n /23g−3+n =
22g−2+n = 2E−V . By comparison one obtains Kontsevich’s main identity with Γg,n the set
of all face-labelled trivalent connected fat–graphs γ of genus g and n faces

X n
Y X
(2df − 1)!! 2−V Y 2
hτd1 . . . τdn i 2df +1
= (1.10)
λf |Autγ | e∈γ λf + λf ′
Σn
1 df =3g−3+n f =1 γ∈Γg,n

where as above V ≡ Vγ , e denote the edges of the graph and 2/(λf + λf ′ ) is the propagator
attached to the edge e bordering f and f ′ .
The right hand side of this expression is suggestive of the Feynman expansion of
the logarithm of the matrix integral (0.1) over N × N Hermitian matrices (N → ∞).
Let Λ stand for a diagonal matrix diag (λ0 , . . . , λN−1 ) and introduce the infinite set t. =
{t0 , t1 , . . .} defined as1
ti (Λ) = −(2i − 1)!! trΛ−2i−1 . (1.11)

1
By convention (−1)!! = 1.

4
By summing the above expression over g and n, one has
X 1
F (t. (Λ)) = hτd . . . τdn i td1 (Λ) . . . tdn (Λ) (1.12)
n! 1
n≥1
d1 ,...,dn ≥0
X  i V 1 Y 2
=
2 |Autγ | e∈γ λf + λf ′
γ∈ΓN

≡ ln Z (N) (Λ)

where the last equality is in terms of the Feynman graph expansion of the integral (0.1)
P
and ΓN refers to the summation over fat graphs with N faces and all possible distinct
assignments of variables λ to their faces. We have noted that (−1)n = iV since the relation
2E = 3V for trivalent graphs implies that the number of vertices V is even = 2p and from
V − E + n = 2p − 3p + n ≡ 0 mod 2, it follows that n is of the parity of p. Formula (0.1)
leads to a propagator 2/(λf + λf ′ ), while the coupling at each vertex is 2i . Thus the above
reads (after letting N → ∞) D P∞ E
ln Z(Λ) = e 1 τd td (1.13)

in agreement with (1.1). In the next section, we discuss the precise mechanism of the
N → ∞ limit by which from finite Λ and finitely many independent trΛ−(2i+1) infinitely
many independent variables t are generated.
The first non trivial graphs in the Feynman expansion yield after some rearrangement
 
t30 t1 t30 t1 1 t21
F = ln Z = + + + t0 t2 + +... (1.14)
3! 24 3! 24 2

2. The Kontsevich integral.

2.1. The main theorem

With the normalized measure

N
Y Y
dM exp − 21 trΛM 2
dµΛ (M ) = R , dM ≡ dMii dℜeMij dℑmMij (2.1)
dM exp − 12 trΛM 2 i=1 i<j

over N × N hermitian matrices M , Λ standing at first for a positive definite such matrix,
we wish to study the matrix Airy function
Z X (N)
(N) i
Z = dµΛ (M ) exp trM 3 = Zk (Λ) (2.2a)
6
k≥0
Z  2k
(N) (−1)k trM 3
Zk (Λ) = dµΛ (M ) (2.2b)
(2k)! 6

5
or rather its asymptotic expansion in traces of powers of Λ−1 for large N . A generalization
is presented in sec. 7.
(N)
More precisely the terms Zk (Λ) can be expressed as polynomials with rational co-
efficients in the variables
1
θr = trΛ−r . (2.3)
r
(N)
Then Zk is homogeneous of degree 3k, if we set

deg θr = r. (2.4)

Only the first N of the θr are algebraically independent for an N × N matrix Λ. However
(N)
Lemma 1 Considered as a function of θ. ≡ {θr }, Zk (Λ) is independent of N for
3k ≤ N and depends only on θr , 1 ≤ r ≤ 3k.
This allows one to define unambiguously the series

X
Z(θ. ) = Zk (θ. ) (2.5)
k≥0

where
(N)
Zk (θ. ) = Zk (θ. ) , N ≥ 3k (2.6)

without any further reference to N . The set of variables θ. is denumerable but each Zk
(Z0 = 1) depends on finitely many of them. This almost evident lemma follows from
eq. (2.43) below.

Theorem 1 (Kontsevich)
∂Z
(i) (θ. ) = 0 r≥1 (2.7)
∂θ2r
(ii) Z(θ. ) is a τ -function for the Korteweg-de Vries equation. Namely if

tr = −(2r + 1)!! θ2r+1 = −(2r − 1)!! trΛ−2r−1

∂2
u = 2 ln Z (2.8)
∂t0
 
∂u ∂ 1 ∂ 2u 1 2
then = + u (2.9a)
∂t1 ∂t0 12 ∂t20 2
∂ ∂
and more generally u= Rn+1 . (2.9b)
∂tn ∂t0

6
In (2.9), the Rn denote the Gelfand-Dikii differential polynomials (derivatives are taken
with respect to t0 )

u2 u′′
R2 = +
2 12
3
u uu′′ u′2 u(4)
R3 = + + +
6 12 24 240
4 ′2 2 ′′
u uu u u uu(4) u′ u′′′ (u′′ )2 u(6)
R4 = + + + + + +
24 24 24 240 120 160 6720
... (2.10)
un
Rn = +···
n!
computed from
′ 1 ′′′
(2n + 1)Rn+1 = R + 2uRn′ + u′ Rn . (2.11)
4 n

Proof of the first part of theorem 1 follows in Kontsevich’s approach from the identities
(1.10)–(1.13), relying on topological considerations, namely on the introduction of the
combinatorial model for Mg,n . The rest of this section is devoted to a purely algebraic
proof of this result.

2.2. Expansion of Z on characters and Schur functions

The integral (2.1) in the case N = 1, denoted for short z(λ),

R∞ 1 2
+ 6i m3   12 Z
−∞
dm e− 2 λm λ ∞
1 2
+ 6i m3
z(λ) = R∞ − 12 λm2
= dm e− 2 λm (2.12)
dm e 2π −∞
−∞

admits the asymptotic expansion

∞
X
z(λ) = ck λ−3k (2.13a)
k=0
 2 k Γ(3k + 12 )
ck = − √ (2.13b)
9 (2k)! π

and satisfies the differential equation equivalent to Airy’s equation

D2 − λ2 z(λ) = 0 (2.14)
1 3 1 1 ∂
− 1 λ3 − 1
D = −e 3 λ λ 2 e 3 λ 2
λ ∂λ
1 1 ∂
=λ+ 2 − . (2.15)
2λ λ ∂λ
7
 For N arbitrary, we can make a shift of variable M → M − iΛ in the numerator of
the integral (2.2.) This gives

Y Y Z  3 
(N) 1 1 Λ3 M M Λ2
Z (Λ) = N (N −1) N2
λr
2
(λr + λs ) exp tr dM exp i tr + ,
2 2 (2π) 2
r r<s
3 6 2
(2.16)
with λi the eigenvalues of Λ. We then integrate over “angles”, i.e. over the unitary group
with the action M → U M U −1 . The result [11] (which we now realize had been established
long before by Harish-Chandra [12]) yields
Z  3 
M M Λ2
dM exp i tr + =
6 2
Z Y
Y m −m 
N2 dmr i 1 3 1 2
6 mr + 2 mr λr
s r
= (2π) 2
1 e λ2s
λ2r
(2.17)
0≤r≤N−1
(2π) 2
0≤r<s≤N−1 − i
i 22
 ∂  
∂  Y − λ3r − 21
N (N −1) N−1
N2 (−1) 2N(N−1)
2 Y
= (2π) 2 Q 2 2
− e 3 λr z(λr ) .
0≤r<s≤N−1 (λs − λr ) ∂λ2s ∂λ2r r=0
0≤r<s≤N−1

with m0 , m1 , · · · , mN−1 the eigenvalues of the matrix M . We denote the Vandermonde

determinant
x00 x10 . . . xN−1
0
x01 x11 . . . xN−1
1 Y
.. .. .. .. = (xs − xr ) (2.18)
. . . . 0≤r<s≤N−1
x0N−1 x1N−1 . . . xN−1
N−1

by |x0 , x1 , . . . , xN−1 |, with the understanding that in each row one substitutes successively
x0 , x1 ,· · ·, xN−1 for the variable x. Inserting (2.17) into (2.16), we find

|D0 z, D1 z, · · · , DN−1 z|
Z (N) (Λ) = , (2.19)
|λ0 , λ1 , · · · , λN−1 |

with D defined as in (2.15). The asymptotic expansion involves only inverse powers of Λ.
We set
1 X
xr = λ−1
r θk = xkr , (2.20)
k
0≤r≤N−1

and consider henceforth the function z as given by the formal series

∞
X
z(x) = ck x3k . (2.21)
0

8
We have D2 z = λ2 z, and one readily sees that

D2k z = λ2k z mod (D2k−1 z, · · · , D0 z)

D2k+1 z = λ2k+1 z mod (D2k z, · · · , D0 z) , (2.22)

with

1
z= Dz (2.23)
λ
   ∞
X
3 1 d
= 1+x +x z= dk x3k
2 dx 0
1 + 6k
dk = ck .
1 − 6k

As a consequence, we rewrite (2.19) as

|λ0 z, λ1 z, λ2 z, · · · |
Z (N) (Λ) = , (2.24)
|λ0 , λ1 , λ2 , · · · , λN−1 |

where the last term in any row of the upper determinant is λN−1 z if N is odd, and
λN−1 z if N is even. Finally we cancel from numerator and denominator the product

N−1
λ0 · · · λN−1 , and express Z (N) in terms of the variables xr = λ−1
r as

(N) |xN−1 z, xN−2 z, · · · |

Z = . (2.25)
|xN−1 , xN−2 , · · · |

Then we substitute the expansions of z and z to obtain the series

X |x3n0 +N−1 , x3n1 +N−2 , · · · , x3nN −1 |

Z (N) = c(0) (1) (N−1)
n0 cn1 . . . cnN −1 , (2.26)
|xN−1 , xN−2 , · · · , x0 |
0≤n0 ,n1 ,···,nN −1

with the convention that

c(2p)
n = cn c(2p+1)
n = dn . (2.27)

When the indices f0 = 3n0 , f1 = 3n1 , · · ·, fN−1 = 3nN−1 form a non-increasing

sequence, we recognize in the above ratio of determinants the symmetric function in
x0 , · · · , xN−1 which corresponds to the polynomial character of the general linear group
GL(N ) specified by the Young tableau with rows of length f0 , f1 , · · ·, fN−1 :

9
 lN−1 = f0 + N − 1

lN−2 = f1 + N − 2
..
.
..
.
..
.

l0 = fN−1
.

The latter admits a natural extension outside the standard Weyl chamber, as an antisym-
metric function of the unordered exponents

lN−1 = f0 + N − 1, · · · l0 = fN−1 . (2.28)

Henceforth we refer to this extension when we write this character as

|xlN −1 , · · · , xl0 |
chlN −1 ,···,l0 = . (2.29)
|xN−1 , · · · , x0 |
PN−1
We wish to express this quantity in terms of the traces θk = 0 xkr . This follows from
the standard identities for characters which we now recall [13]. Thinking of X as the
diagonal matrix X ≡ Λ−1 = diag (x0 , x1 , · · · , xN−1 ) we set s0 = p0 = 1 and for k ≥ 1

N
X
sk (X) =tr ∧k X det(1 + uX) = uk sk (X) (2.30a)
0
∞
X
1
pk (X) =tr ⊗ksym X = uk pk (X) (2.30b)
det(1 − uX) 0
1
θk = trX k . (2.30c)
k

The quantities sk , 1 ≤ k ≤ N , are the elementary symmetric functions corresponding to

the vertical Young tableaux up to N lines, whereas the pk ’s are the traces of symmetric
tensor products, i.e. they correspond to Young tableaux with only one row.
We have from the definition
∞
X ∞
X
n
exp u θn (X) = uk pk (X) , (2.31)
1 0

10
which expresses the pk ’s as homogeneous polynomials of degree k of the θn ’s (deg θn = n),
ignoring the relations among traces. This justifies the definition of (elementary) Schur
functions pn via
∞
X ∞
X
n
exp u θn = uk pk (θ. ) , (2.32)
1 0

without reference to any N × N matrix, and where now the p’s are functions of the θ’s.
When both p and θ refer to the same matrix X we recover the previous definitions (2.30)
and (2.31). Explicitly we write
X θ1ν1 θ2ν2
pr (θ. ) = ... . (2.33)
ν1 +2ν2 +...=r
ν1 ! ν2 !

Eq.(2.32) entails
∂pr (θ. ) ∂ k pr (θ. )
= = pr−k (θ. ) , (2.34)
∂θk ∂θ1k
where pn vanishes for n < 0. When expanding the matrix elements along successive
columns, Cauchy’s determinental formula

1 |xN−1 , · · · , x0 | |y N−1 , · · · , y 0 |
det = Q (2.35)
1 − xr y s 0≤r,s≤N−1 r,s (1 − xr ys )

yields
X l |xlN −1 , · · · , xl0 | |y N−1 , · · · , y 0 |
l0
y0N −1 · · · yN−1 = Q . (2.36)
|xN−1 , · · · , x0 | r,s (1 − xr ys )
l0 ,···,lN −1

Therefore if X ≡ diag ({xr }), then

y0N −1 y00
det(1−y0 X) ... det(1−y0 X)
l l0
chlN −1 ,···,l0 (X) = coeff. of y0N −1 · · · yN−1 in .. .. .. . (2.37)
. . .
N −1 0
yN −1 yN −1
det(1−yN −1 X)
... det(1−yN −1 X)

h i−1
Expanding det(1 − yX) according to (2.30b), we obtain the classical formula (Jacobi-
Schur)

plN −1 −(N−1) (X) . . . plN −1 −1 (X) plN −1 (X)

|xlN −1 , · · · , xl0 | plN −2 −(N−1) (X) . . . plN −2 −1 (X) plN −2 (X)
chlN −1 ,···,l0 (X) = N−1 = .. .. ,
|x , · · · , x0 | . .
pl0 −(N−1) (X) . . . pl0 −1 (X) pl0 (X)
(2.38)

11
valid for any ordered or unordered sequence lN−1 , · · · , l0 . Terms along the diagonal read
pf0 , pf1 , · · · , pfN −1 , and indices increase (decrease) by successive units as one moves from
a diagonal term to the right (left). We abbreviate this expression as
pf0 ⋆ ... ⋆
⋆ pf1 ... ⋆
chN−1+f0 ,···,fN −1 (θ. ) = . .. .. .. , (2.39)
.. . . .
⋆ ⋆ . . . pfN −1
substituting for the elementary Schur polynomials their expressions in terms of the vari-
ables θ. . We conclude that
p3n0
X
Z (N) = c(0) (1) (N−1) ..
n0 cn1 · · · cnN −1 . (2.40)
n0 ,···,nN −1
p3nN −1
yields an expression of Z (N) in terms of the infinitely many variables θ. (which can hence-
forth be treated as independent). It follows from eq. (2.40) that
p3n0
X
(N)
Zk = c(0) (1)
· · · cn(N−1) ..
n0 cn1 N −1 . (2.41)
n0 +···+nN −1 =k p3nN −1
where each character is of degree 3k. This is obvious for the diagonal term and, as one
readily ascertains, holds also for non-diagonal terms.
(N)
We are now in position to prove the lemma, which is trivially true for Z0 = 1.
Suppose 0 < 3k ≤ N and for a given term in (2.41) let δ be its “depth”, i.e. the smallest
integer ≤ N such that r ≥ δ ⇒ nr = 0. From (2.39) it follows that the corresponding term
reads
p3n0
c(0) (1) (δ−1) ..
n0 cn1 · · · cnδ−1 . ; n0 + · · · + nδ−1 = k.
p3nδ−1

T
The last column of the δ×δ determinant reads p3n0 +δ−1 , · · · , p3nδ−1 where the subscripts
Pδ−1
are δ positive integers with a sum equal to 3k + 0 r. If 3k < δ, this is smaller than the
sum of the first δ positive integers. From Dirichlet’s box principle, two of the subscripts
among 3n0 + δ − 1, 3n1 + δ − 2, · · ·, 3nδ−1 have to be equal, which results in two identical
lines in the determinant which therefore vanishes. Hence δ has to be smaller than or equal
to 3k, showing that
(N) (3k)
N ≥ 3k =⇒ Zk = Zk ≡ Zk (2.42)
which concludes the proof and allows a definition of the formal series Z, without reference
to N .

12
2.3. Proof of the first part of the Theorem.

(i) The first part of the theorem will be proved for each Zk which we take equal
(3k)
to Zk . Differentiating each line successively in the determinental characters using the
crucial formula (2.34) we find

∂Zk
2r > 3k =0 (2.43)
∂θ2r
 ∂Z P3k−1

 ∂θ
k
= s=0 Zk,(s)


2r

 p3n0


 ..
.
2r ≤ 3k P (0) (1) (3k−1)

 Zk,(s) = c c
n0 +···+n3k−1 =k n0 n1 · · · cn3k−1 p3ns −2r



 ..

 .

p3n3k−1

where subscripts in the s-th row of each determinant have been decreased by 2r units. For
0 ≤ s ≤ 3k − 1 − 2r the subscripts in line s and s + 2r only differ by the interchange of
ns and ns+2r . The determinental character is therefore antisymmetric in the interchange
(s) (s+2r)
of indices ns and ns+2r whereas in the product of c’s, due to (2.27) . . . cns . . . cns+2r . . .
(s) (s)
≡ . . . cns . . . cns+2r . . . is symmetric in these indices. As a consequence, Zk,(s) vanishes for
s ≤ 3k − 1 − 2r and we need only consider terms with s > 3k − 1 − 2r, i.e. when the
derivative acts on one of the last 2r lines and we cannot use the above argument relying
(r) (r+2)
on the periodicity cn = cn .
(ii) Therefore fix s such that 3k − 2r ≤ s ≤ 3k − 1. The only possibly non-vanishing
terms in Zk,(s) are those whose depth δ defined as above to be the smallest integer such
that r ≥ δ ⇒ nr = 0, satisfies the inequality 3k − 2r ≤ s ≤ δ − 1 ≤ 3k − 1. In this case
they read

p3n0
..
.
c(0) (1)
n0 cn1 · · · c(δ−1)
nδ−1 p3ns −2r , nδ−1 > 0, n0 + . . . + nδ−1 = k . (2.44)
..
.
p3nδ−1

The δ indices of the p’s in the last column of the determinant before subtracting 2r from
Pδ
the indices of the s-th row are all positive integers and have a sum equal to 3k − δ + 1 t,
where 0 ≤ 3k − δ ≤ 2r − 1. We now make use of the following

13
Lemma 2.
If from a set of δ positive integers, with sum exceeding the one of the first δ positive integers
by an amount ∆ ≥ 0, one decreases one by ∆′ > ∆, then in the new sequence two terms
coincide or one is a non positive integer.
Think of the original set as occupied integral levels. Let r0 + 1 be the first unoccupied
one (r0 ≥ 0) and r1 the greatest occupied one. It follows from the hypothesis that r1 −r0 ≤
∆ + 1. If one decreases one element of the set by an amount ∆′ ≥ ∆ + 1, it therefore
becomes less than or equal to r0 , which proves the lemma.
Applying this result to the above circumstance (∆ = 2r − 1, ∆′ = 2r), we deduce
that the only possibly non-vanishing terms in Zk,(s) , 3k − 2r ≤ s ≤ 3k − 1 occur when
3ns = 2r − (δ − 1 − s) with 0 ≤ δ − 1 − s ≤ 2r − 1. Thus 2r − (δ − 1 − s) takes the possible
values 1, . . . , 2r. If r = 1 this is never a multiple of 3. Thus we have already obtained

∂Z
= 0. (2.45)
∂θ2

We can even say more. Let a be the integral part of (δ − 1)/3 and consider in the last
column starting from the bottom the a + 1 positive subscripts

3nδ−1 , 3(nδ−1−3 + 1), . . . , 3(nδ−1−3a + a) .

For the corresponding character to be non-zero, these have to be all distinct. Hence their
Pa
sum is larger than or equal to 3 0 (α + 1). The inequality

a
X a+1
X


3 nδ−1−3α + α ≥ 3 α (2.46)
α=0 0

Pa
implies that α=0 nδ−1−3α ≥ a + 1. Should a non-vanishing term arise in Zk,(s) , there
would exist an index ns such that from the preceding observation

3ns + δ − 1 − s = 2r (2.47)

with
3k − 2r ≤ s ≤ δ − 1 . (2.48)

Thus 3ns + δ − 1 = 2r + s ≥ 3k, or equivalently ns + a ≥ k. If 2r is not a multiple of 3, i.e.

r 6= 0 mod 3 , (2.49)

14
it follows that
δ − 1 − s 6= 0 mod 3 . (2.50)

This means that the subscript 3ns +δ−1−s does not belong to the sequence 3nδ−1−3α +3α.
Since the sum of all n’s is k we should have
a
X
k≥ nδ−1−3α + ns ≥ a + 1 + ns (2.51)
0

whereas from the above a + 1 + ns ≥ k + 1, a contradiction. Thus in general

∂Z
r 6= 0 mod 3 = 0. (2.52)
∂θ2r
(iii) The remaining cases are those for which we take a derivative with respect to θ6r .
We shall need a relation between the coefficients cn , dn which did not play any specific
role until now. The series z(x) of (2.21) is a solution of the following differential equation
in the variable x
5
x4 z ′′ + 2(2x3 + 1)z ′ + x2 z = 0 , (2.53)
4
while z(x) of (2.23) satisfies
∞
X  
3n x3
z(x) = dn x = 1+ z + x4 z ′ . (2.54)
0
2

Lemma 3
z(x)z(−x) + z(−x)z(x) = 2 (2.55)

To see this, substitute z in terms of z and compute the derivative with respect to x,
making use of the differential equation for both z(x) and z(−x), to find that it vanishes.
In terms of the series expansions this reads
2n
X
(−1)s c2n−s ds = 0 n > 0. (2.56)
s=0

It was convenient to use generalized characters until now, but we can also recast the
expansion of Z in terms of standard characters indexed by Young tableaux Y (f0 ≥ f1 ≥
. . . ≥ fδ−1 > 0) at the price of having more complicated coefficients. From eqs. (2.38) and
(2.40) this reads
X
Z= ξY chY (θ. ) (2.57)
Y, |Y |=0,mod 3

15
with
chY (θ. ) = det{pfi +j−i (θ. )} 0 ≤ i, j ≤ δ − 1 (2.58)

and

ξY = det ξi,j (2.59)

(
0 if fi + j − i 6= 0 mod 3
ξi,j = c(j) if fi + j − i = 0 mod 3 .
1
(f +j−i)
3 i

In the above, i (j) is the row (column) index. The diagonal subscripts in chY are no longer
multiples of 3, instead running down the diagonal they form a non-decreasing sequence,
the last one being positive (and δ being the number of rows of the corresponding Young
tableau). The quantity Zk is obtained by restricting the sum to |Y | = 3k.
When taking a derivative with respect to θ6r , the first contributing term is Z2r .
(Recall that terms in Zk are homogeneous of degree 3k). Let us therefore investigate first
∂Z2r
∂θ6r . The only Young tableaux such that |Y | = 6r for which ∂chY /∂θ6r 6= 0 are of the
“Fermi-Bose” type (t + 1)(1)s , s + t + 1 = 6r

t+1
z }| {







s (2.60)






in which case ∂chY /∂θ6r = (−1)s . Applying formula (2.57) we get

(s)
c2r−s d2r−s+1 . . . c2r
2r−1
X (s)
∂Z2r 1 d1 . . . cs
=2 (−1)s .. .. ..
∂θ6r s=0 . . .
(s)
0 ... 1 c1
(s+1)
d2r−s c2r−s+1 . . . c2r
2r−1
X (s+1)
1 c1 . . . cs
+ (−1)s .. .. .. . (2.61)
s=0 . . .
(s+1)
0 ... 1 c1

16
Expanding each of these determinants along the first column (where c and d with negative
index are set equal to zero), we find

r
∂Z2r X  
= (1 − 3j)c2j−1 ∆2r−(2j−1) + 3jc2j ∆2r−2j
∂θ6r j=0
r
X  
+ 3jd2j ∆2r−2j − (3j − 2)d2j+1 ∆2r−2j−1 , (2.62)
j=1

where ∆0 = ∆0 = 1 and
(s) (s+1)
d1 c2 . . . cs c1 d2 . . . cs
(s) (s+1)
1 c1 . . . cs−1 1 d1 . . . cs−1
∆s = . .. .. ∆s = . .. .. . (2.63)
.. . . .. . .
(s) (s+1)
0 ... 1 c1 0 . . . 1 c1

Taking into account the identity (2.56) satisfied by the coefficients c and d we readily see
by a recursive argument that ∆ and ∆ reduce to

∆s = d s ∆s = cs (2.64)

Thus
X2r
∂Z2r
= (3r − 1) (−1)j cj d2r−j = 0. (2.65)
∂θ6r j=0

(iv) It remains finally to examine

∂Z2r+k
k>0
∂θ6r

Let us look at the expression (2.43) with the required modification k → k +2r, 2r → 6r. As
before when taking derivatives of the row of index s we need only take into account those
terms for which s ≥ 3k, the others vanishing due to the antisymmetry of the characters.
This being assumed we consider for fixed s a specific term in the sum (2.43) with character
of depth δ > 3k so that we can omit the rows and columns of label larger than δ − 1
in the computation of the corresponding determinental character. The labels in the last
column before derivation are 3n0 + δ − 1, . . .,3nδ−1 , no pair of them equal. According to a
previous analysis using lemma 2, to get a non-trivially vanishing derivative, the quantity
3ns +δ−1−s (s ≥ 3k) has to equal 6r . This means that δ−1−s = 3σ, and nδ−1−3σ = 2r−σ.

17
 Pσ−1
Since by definition nδ−1 > 0, from the preceding reasoning we must have α=0 nδ−1−3α ≥
σ, if this sum is non-empty (i.e. σ > 0). Then
σ
X
nδ−1−3α ≥ 2r , (2.66)
α=0

an inequality which remains obviously true when σ = 0, in which case nδ−1 = 2r. A
fortiori
δ−1
X
nρ ≥ 2r , (2.67)
ρ=3k
Pδ−1
since the latter sum includes the previous one (δ − 1 − 3σ ≥ 3k). Since 0 nρ = 2r + k
from the homogeneity property of Z2r+k , we have the complementary inequality
3k−1
X
nρ ≤ k. (2.68)
0

Both inequalities must in fact be equalities. Indeed among the integers of the form δ−1−3α
(α ≥ 0) such that δ − 1 − 3α ≥ k, consider the largest one, say β, for which nδ−1−3α > 0.
We have β ≥ σ and again appealing to a previous reasoning
X
nρ ≥ k , (2.69)
0≤ρ<δ−1−3β
ρ=δ−1 mod 3

since there are at least k integers equal to δ − 1 mod 3 between 0 and 3k − 1 < δ − 1 − 3β.
Hence
Pσ
(i) α=0 nδ−1−3α has to equal 2r otherwise we would violate homogeneity;
(ii) β has to be equal to σ for the same reason (no other nδ−1−3α except those entering
the previous sum are > 0);
P
(iii) and finally should nρ0 , ρ0 ≥ 3k, ρ0 6= δ−1 mod 3, be positive, then again 0≤ρ≤ρ0 nρ
ρ=ρ0 mod 3
would be larger than k (since the sum includes ρ0 ≥ 3k), a contradiction. We conclude
that we can replace (2.68) and (2.69) by equalities. This means that we can write

p3n0
..
X X .
∂Z2r+k
= c(0)
n0 · · · c(6r+3k−1)
n6r+3k−1 p3ns −6r
∂θ6r ..
s≥3k n0 +···+n3k−1 =k
n3k +...+n6r+3k =2r .
p3n3k+6r−1
(2.70)

18
Let us group together all contributions corresponding to a fixed choice of indices
n0 , . . . , n3k−1 with sum equal to k. Taking into account the antisymmetry in the last
6r rows we see that the analysis reduces to our previous computation of ∂Z2r /∂θ6r . To be
precise, the column labelled 3k will correspond to coefficients labelled c or d according to
k even or odd so that if k is even

p3n0
∂Z2r+k X ∂Z2r
= c(0) · · · c(3k−1) .. . (2.71)
∂θ6r n0 n3k−1 . ∂θ6r
n0 +···+n3k−1 =k p3n3k−1

If k is odd, we get the same formula with ∂Z2r /∂θ6r replaced by the same expression with
c’s and d’s interchanged. In both cases the expression vanishes since c’s and d’s play a
symmetric role in the vanishing of ∂Z2r /∂θ6r as it relies on the identity (2.55) invariant in
the interchange of z and z.
We have at last fully proved the first part of Kontsevich’s theorem from a purely
algebraic standpoint
∂Zk
=0 (2.72)
∂θ2r
Even though the proof appears a little long, the steps are completely elementary relying on
the second order differential equation satisfied by z and Weyl antisymmetry of characters.
Without repeating in detail each step, it will go through in the generalized case considered
in sec. 7.
It may also be worth remarking that by retracing the above discussion, this property
is very likely to imply the (Airy-like) differential equation. We now turn to the second
part of the theorem.

3. From Grassmannians to KdV

Expert readers will have recognized the connection between the expansion (2.57) of Z and
Sato’s approach to soliton equations and τ - functions. The latter relies on a clever rein-
terpretation of the familiar Plücker relations of projective geometry in terms of properties
of associated (pseudo-)differential operators [14]. In more physical terms, this involves the
characterization of those submanifolds that correspond to pure Slater determinants (as
opposed to their linear combinations) in a many body fermionic space.
Let us begin with a short review of the subject following Sato. Let V be a vector space
of dimension N equipped with a basis e0 , · · · , eN−1 . The field of constants is arbitrary but

19
one may think of IR or C. A linear subspace generated by m vectors ξ (0) , · · · , ξ (m−1) is
intrinsically described by the antisymmetric multivector
X
ξ (0) ∧ · · · ∧ ξ (m−1) = ξl0 ,···,lm−1 el0 ∧ · · · ∧ elm−1 (3.1)
0≤l0 ≤···≤lm−1 ≤N−1

with antisymmetric components

ξl0 ,···,lm−1 = det ξli ,j , 0 ≤ i, j ≤ m − 1 , ξ (k) = ξi,k ei (3.2)

All relations to be written being homogeneous we may consider the above quantities as
being homogeneous components of the corresponding m − 1 dimensional linear subspace
in the projective space P V (N − 1) of dimension N − 1. A familiar case is the description
of lines (m = 2) ¡in projective three space (N = 4). An (m − 1)-dimensional subspace in
P V (N − 1) depends on m(N − m) parameters2 (four in the above example, for instance
the intersections of the line with two planes) while the number of coordinates in (3.1)


(taking into account homogeneity) is Nm − 1 (i.e. 5 in the example). They must therefore
satisfy some (non-linear but homogeneous) relations. These are the Plücker relations. In
the aforementioned example this is the classical quadratic relation expressing that the
geometry of lines in the projective 3–dimensional space is equivalent to the geometry of
points on a quadric in 5–dimensional projective space.
To derive typical Plücker relations, we demand that a linear combination
X
η= xk ξ (k) (3.3)

lies in the subspace generated by the ξ’s, i.e. that

X
0 = η ∧ ξ (0) ∧ · · · ξ (m−1) = ξel0 ,···,lm el0 ∧ · · · ∧ elm (3.4)
0≤l0 ···lm ≤N−1
m
X m−1
X (r)
ξel0 ,···,lm = (−1)i xr ξli ξl b .
0 ,···,li ,···,lm
i=0 r=0

Choose the coefficients xr as the minors in the last line of

(0) (m−1)
ξk0 . . . ξk0
.. .. m−1
X
ξk0 ,···,km−2 ,l ≡ . . = xr ξ l
(r)
(3.5)
(0) (m−1)
ξkm−2 ... ξkm−2 r=0
(0) (m−1)
ξl ... ξl
2
This is m × N , the number of components of the vectors ξ (k) , minus m2 , the dimension of
the linear group GL(m) acting linearly on this vector without modifying the subspace.

20
which only depend on the choice of k0 , · · · , km−2 and not on the index l. Hence we get the
Plücker relations in the form
m
X
ξk ξl b = 0. (3.6)
0 ,···,km−2 ,li 0 ,···,li ,···,lm
i=0

The reader might have fun to find the relations among these relations and so on. In
any case by turning the argument around these relations do characterize m–dimensional
vector subspaces in V or the (m − 1)–dimensional ones in P V which form the (N, m)
Grassmannian (obviously not a vector space but rather an intersection of quadrics).
For our purposes we will need a generalization of (3.5) which follows from the obser-
vation that we could equally well replace η by some combination
X ′
η′ = xk,k′ ξ (k) ∧ ξ (k ) (3.7)
k<k′

(or for that matter by any higher superposition of exterior products). Leaving the general
case aside, we also have
X
(−1)i+j−1 ξk0 ,···,km−3 ,li ,lj ξl b b (3.8)
0 ,...,li ,···,lj ,...,lm+1
i<j

and so on.
The relevance of this discussion to the present problem arises from the determinen-
tal expressions for the function Z as expressed in eqs. (2.26) and (2.57). Indeed it was
Sato’s idea to associate to points of the Grassmannian a τ –function obtained by replacing
exterior products of basis vectors by the corresponding antisymmetric generalized Schur
functions. In our case one can view vector subspaces as those generated by the functions
z, Dz, D2 z, · · · so that Kontsevich’s integral appears as a realization of Sato’s idea. The
task is now to translate equivalents of the Plücker relations in terms of Z.
It will be easier to state this in the finite N case we started from, since by letting N
become arbitrarily large we will recover the required results term by term in the asymptotic
series. Thus we return to formula (2.40) understanding by pn (θ. ) the unconstrained Schur
functions which we recast in the following form
∂ N −1 fN −1 ∂fN −1
∂θ1N −1
... ∂θ1
fN−1
Z (N) = .. .. , (3.9)
.
N −1
.
∂ f0 ∂f0
∂θ1N −1
... ∂θ1
f0

21
where
X
fN−1 (θ. ) = c(0)
n p3n+N−1 (θ. )
X
f1 (θ. ) = c(1)
n p3n+N−2 (θ. )

..
. (3.10)
X
f0 (θ. ) = cn(N−1) p3n (θ. ) .

In the above we have used eq. (2.34) which is meaningful for Schur functions with in-
dependent θ arguments. The precise meaning of (3.10) for unconstrained θ’s is therefore
that it extracts from the complete formula (2.57) only those terms corresponding to Young
tableaux which have at most N rows. By letting N → ∞ we reach any desired term.
The expression (3.9) singles out the variable θ1 , the others playing the role of param-
eters. For the time being we simplify the notations by referring to Z (N) as Z until at the
end we restore the correct subscript. Looking at (3.9) we see that it takes the form of a
Wronskian of the components f0 , . . . , fN−1 of a vector denoted f . It is therefore natural
to attach it to an N –th order differential operator ∆N such that for an arbitrary function
d
F of θ1 (d ≡ dθ1 )

∂N F
∂θ1N
... F
N
N ∂ fN −1
X . . . fN−1
∂θ1N
∆N F = wr (θ. ) dN−r F = Z −1 (3.11)
.. ..
r=0 . .
N
∂ f0
∂θ1N
... f0

Strictly speaking the coefficients w should also carry the subscript N . In order to obtain
a smooth transcription as N → ∞, Sato uses rather than the differential operator ∆N an
equivalent pseudodifferential operator W defined as

∆N = WN dN , (3.12)
N
X
WN = wr d−r (3.13)
0

with w0 = 1 and
∂

w1 = Z −1 − Z. (3.14)
∂θ1
Expanded in power series in the θ’s, w1 will have terms of fixed degree independent of
N for N large enough and the remark applies to the successive coefficients wp (infinitely

22
many as N → ∞) justifying that we drop eventually all reference to N . Equation (3.14)
admits a rather neat generalization as
 
1 ∂
wr = pr − Z , (3.15)
Z ∂θ.

where
∂ ∂ 1 ∂ 1 ∂
≡ , ,..., ,... (3.16)
∂θ. ∂θ1 2 ∂θ2 k ∂θk
(which is meaningful since we consider the θ’s as independent variables). Indeed let us
form the generating function
X    
∂ 
∞
X
−1 r −1 yr ∂
Z y pr − Z=Z exp − Z. (3.17)
∂θ. 1
r ∂θr
r≥0

∂ ∂r
Acting on any column vector of Z, the operator ∂θr is equivalent to ∂θ1r . This gives
X  
−1 r ∂ 
Z y pr − Z = Z −1 det (1 − yd)dN−1 f , . . . , (1 − yd)f . (3.18)
∂θ.
r≥0

Let us compare this with the quantity

N
X y0 . . . yN
wr y r = Z −1 det . (3.19)
dN f ... f
0

The (N + 1) × (N + 1) determinant on the right hand side can be computed by subtracting

the first column multiplied by y from the second, the second column multipled by y from
the third and so on, with the cofactor of the only non-vanishing element in the first row
equal to the previous expression proving eq. (3.15).
One is familiar with the commutation relations involving the operators d−r
X  
−r k r+k−1
d a= (−1) a(k) d−r−k
k
k≥0
X r + k − 1
−r
ad = d−r−k a(k) (3.20)
k
r≥0

where a(k) ≡ (dk a). These formulae enable one to give a meaning, again droping the index
N , to the “dual” pseudo–differential operator
X  
∗ −r ∂
W = d wr∗ wr∗ =Z −1
pr Z (3.21)
∂θ.

23
satisfying
W ∗ = W −1 (3.22)

We relegate the (cumbersome) proof of this latter fact which relies on Plücker formulae to
an appendix of this section.
The crux of the matter is the basic equation

∂W
= Qn W − W dn (3.23)
∂θn

with Qn a normalized differential operator of order n, Qn = dn + . . ., given by

 
n −1
Qn = W d W (3.24)
+

the subscript + refers to the differential part of W dn W −1 , as follows from the fact that
∂W
∂θn is of order d−1 .
Multiplying it on the right by dN , eq. (3.23) is equivalent to
 
∂∆N −1
= Qn ∆N − ∆N dn Qn = ∆N d n
∆N (3.25)
∂θn +

where ∆−1
N =d
−N
W −1 . Again N is assumed much larger than n fixed, in fact as large as
we want. Among these equations one is trivially verified namely the one for n = 1 where
Q1 = d. To prove (3.25) choose Qn as indicated so that the combination Qn ∆N − ∆N dn
is a differential operator of order N − 1 as it should be if the formula is to make sense.
It will then hold if both sides agree when operating on N linearly independent functions
which (see eq. (3.11)) we naturally choose as being f0 , . . . , fN−1 or equivalently on any
linear combination which we denote by f . Thus ∆N f = 0 and we want to prove that
 ∂∆  dn f
N
f + ∆N n = 0. (3.26)
∂θn dθ1

dn f ∂f
But dθ1n
through the definition (3.10) is equal to ∂θ reducing the above expression to
∂

n

∂θn ∆N f = 0. Thus eqs. (3.25) and (3.24) hold true and yield in general the so-called
∂2W ∂2W
KP hierarchy of compatible integrable systems ∂θn1 ∂θn2
= ∂θn2 ∂θn1
for the coefficients of
the pseudo-differential operator W . An equivalent form is as follows. Set

L = W dW −1 = d + O(d−1 )


Ln = W dn W −1 Qn = Ln +
. (3.27)

24
Then from (3.23) we have

∂W −1
= − W −1 Qn + dn W −1
∂θn
∂L
=[Qn , L] , (3.28)
∂θn
where in writing these formulae we have implicitly taken the N → ∞ limit. As a conse-
quence of (3.28) we have the zero curvature conditions

∂Qm ∂Qn
− + [Qm , Qn ] = 0. (3.29)
∂θn ∂θm

According to equations (3.15) and (3.21) the coefficients in W and W −1 do not depend on


θ2r . Hence the differential operators Q2r ≡ L2r + commute with L as well as any of its
powers
[Q2r , Q2r′ ] = 0. (3.30)
∂2 ∂2
Using the notations of theorem 1, tr = −(2r + 1)!! θ2r+1 , u = ∂θ12
ln Z = ∂t20
ln Z and
(3.15), (3.21) and (3.24)

Q2 =d2 + 2u
3 ∂u 3

Q3 =d3 + 3ud + ≡ Q22 + (3.31)
2 ∂θ1
Setting m = 2 and n = 3 in (3.29) we conclude that the first non trivial equation in the
hierarchy reads
∂Q2
= [Q3 , Q2 ] (3.32)
∂θ3
i.e.  
∂u ∂ 1 ∂ 2u 1 2
= + u (3.33)
∂t1 ∂t0 12 ∂t20 2
∂
as claimed in the second part of Theorem 1. Higher equations involve ∂t2 , · · · and are of
the form (2.9b),(2.10).
In fact, the commutation of L and Q2 implies that L2 = Q2 , i.e. that L2 is a differential
operator. To prove this, one may appeal to a lemma [15] that asserts that the space
of operators that commute with Q2 is spanned by the powers of the pseudodifferential
operator square root of Q2 , with constant coefficients. Thus

1
∞
X  1 −l
L = Q2 +
2
αl Q22 . (3.34)
l=0

25
 1
Both L and Q22 , however, are functionals of Z with the limit d as Z → 1. It follows that
all the constants α vanish and
1
L = Q22 . (3.35)

This implies that the KdV flows (3.28) are generated by the
r+ 21
Q2r+1 = (L2r+1 )+ = (Q2 )+ .

Notice that the above identity (3.35) means that L, which a priori depends on all the
derivatives of Z, is actually a functional of the sole u = ∂ 2 ln Z/∂θ12 .

Appendix
It would seem that Plücker relations have not entered directly the discussion. One place
where they play a hidden role is in the computation of the inverse W −1 = W ∗ . Of course
if we need only the first few terms as in (3.31) one can obtain them by a direct calculation.
For completeness we give a recursive proof of eq. (3.22). The integer N being fixed we
start with the expressions (3.9)–(3.11) and consider f in (3.10) as a column vector function
∂f
of independent θ’s with d ≡ ∂
∂θ1 and f (r) ≡ ∂θ1r . Dropping the index N

∆ =W dN
N
X
W = wr d−r (3.36)
0
 
∂ d
wr =Z −1
pr − Z = Z −1 (−1)r f (N) . . . f (N−r) . . . f
∂θ.
using a shorthand notation for determinants. The kernel of ∆ is the finite di-
mensional vector space generated by the components of f . We distinguish a flag
(f0 ), (f0 , f1 ), (f0 , f1 , f2 ), · · · and associate to it a sequence of determinants
f1′ f1
Z (1) = f0 Z (2) = , · · · , Z (N) ≡ Z (3.37)
f0′ f0
in terms of which one can write a factorized form (the Miura transformation)
 Z (N) Z (N−1)   Z (2) Z (1)  1 
(1)
∆= d · · · d Z d . (3.38)
Z (N−1) Z (N) Z (1) Z (2) Z (1)
It is clear that applied to Z (1) = f0 , ∆ gives 0 while if ∆k is the product of the first k
factors starting from the right and if we assume ∆k f0 = ∆k f1 = . . . = ∆k fk−1 = 0, then
Z (k+1)
∆k f k = Z (k)
, hence
 Z (k+1) Z (k) 
∆k+1 fk = d (k+1) fk = 0
Z (k) Z
26
proving the above factorization. The identity to be established is therefore

 1  Z (2) −1 Z (1)   Z (N) (N−1) 

−1 Z
W −1 = dN ∆−1 = dN Z (1) d−1 (1) d · · · d
X Z Z (1) Z (2) Z (N−1) Z (N)
= d−r wr∗ (3.39)
r≥0

with coefficients wr∗ given by  

∂
wr∗ =Z −1
pr Z . (3.40)
∂θ.
Upon taking a generating function

X X
∞
yn ∂  1 1
y r
wr∗ =Z −1
exp Z = Z −1 ∂
f (N−1) . . . f (3.41)
r≥0 1
n ∂θn 1 − y ∂θ1 1 − y ∂θ∂ 1

where upper indices on f label derivatives. We have recognized that acting on each column
of Z the shift operator is equivalent to
∞
X  n
yn ∂ 1
exp = .
1
n ∂θ1 1 − y ∂θ∂ 1

If in the above determinant we subtract from the last column the preceding one multiplied
by y and so on, we get
wr∗ = Z −1 |f (N−1+r) , f (N−2) · · · f |. (3.42)

Comparing (3.36) and (3.39), we see that the determinental numerators have a natural
pictorial description in terms of Young tableaux. The first determinant is a vertical Young
tableau, the second a horizontal one. This parallels the correspondence between pr (−θ. ) =
(−1)r sr (θ. ) (recall (2.30)) and pr (θ), and the formula to be established is similar to the
identity det(1 − X) det(1 − X)−1 = 1.
In any case with this expression for wr∗ we return to (3.39) and note that (3.42) says
that w0∗ = 1 in agreement with (3.39) for every N whereas, should N = 1, W −1 reduces to

X −r f0(r)
d f0 d−1 f0−1 = d (3.43)
f0
r≥0

again in agreement with (3.42). We therefore assume that (3.42) holds for any r if N ′ < N
and for r ′ ≤ r when the size of determinants is N , and establish it for wN,r+1
∗
reinstating

27
the index N . We have
Z (N) −1 Z (N−1)
−1
WN−1 = dWN−1 d
Z (N−1) Z (N)
X Z (N) −1 Z (N−1)
= d1−k wN−1,k
∗
(N−1)
d (3.44)
k≥0
Z Z (N)
X X  Z (N) (l) Z (N−1)
= d−r ∗
wN−1,k .
r≥0
Z (N−1) Z (N)
k+l=r

Write
∗ vN,r
wN,r = . (3.45)
Z (N)
We have
X  Z (N)
(l)
(N−1)
vN,r = Z vN−1,k (N−1) 2 . (3.46)
Z
k+l=r

We want to show that

vN,ρ = |f (N−1+ρ) , f (N−2) . . . f | (3.47)

assuming it to be true for N ′ < N (where we only keep the components f0 , . . . , fN ′ −1 ) and
also for ρ ≤ r to prove that it holds for r + 1. Take a derivative of the above identity

Z (N−1)′ X  Z (N)
(l+1)
′ (N−1)
vN,r = (N−1) vN,r + Z vN−1,k (N−1) 2 . (3.48)
Z k+l=r
Z

(N )
The last sum differs from vN,r+1 by the missing term vN−1,r+1 ZZ(N −1) . Hence

′ Z (N) Z (N−1)′
vN,r+1 = vN,r + vN−1,r+1 vN,r . (3.49)
Z (N−1) Z (N−1)

T
Denote the column vector f0 , · · · , fN−2 by ϕ (of dimension N − 1). According to the
recursive hypothesis this reads
γ
vN,r+1 = |f (N+r) , f (N−2) , · · · , f | + (3.50)
Z (N−1)
with

d , ϕ(N−1) d
d , ϕ(N−2) . . . , ϕ f (N−1+r) , f (N−1) , f (N−2)
γ = ϕ(N−1+r ,···,f
d , ϕ(N−1) , ϕ(N−2) d
d . . . , ϕ f (N−1+r) , f (N−1)
− ϕ(N−1+r , f (N−2) , · · · , f
d
d . . . , ϕ f (N−1+r)
d , ϕ(N−2)
+ ϕ(N−1+r , ϕ(N−1) , f (N−1) , f (N−2) , · · · , f (3.51)

28
where in each term the first determinant is (N − 1) × (N − 1) dimensional, the second
N × N . We have to show that the combination γ vanishes since we wish to prove that
vN,r+1 is the first term of the r.h.s. of (3.50). One easily checks that γ = 0 if N = 2, so
we henceforth assume N > 2. To reduce the vanishing of γ to one of Plücker ’s identities,
expand the N × N determinants involving f and its derivatives according to its first line.
(k)
For each term of the form fN−1 the coefficient is a combination of ϕ–determinants which
vanishes by virtue of the Plücker relations (3.8), completing the proof of formulas (3.21)
and (3.22).

4. Matrix Airy equation and Virasoro highest weight conditions.

The differential equation (2.14) generalizes to the N –dimensional case as follows. Call
Y (Λ) the integral appearing in (2.16) for finite N
Z  
M3 M Λ2
Y (Λ) = dM exp itr + . (4.1)
6 2

The function Y satisfies for each index k

Z
d i
0= dM exp (3trM Λ2 + M 3 ) (4.2)
dMkk 6

i.e.
D X E
2
0= Mkl Mlk + Mkk + λ2k (4.3)
l, l6=k
 
M3 M Λ2
where h.i denotes an integral taken with respect to the weight dM exp itr 6
+ 2
.
The insertion of a diagonal factor Mkk can be achieved by acting with the derivative
operator −i λ1k ∂λ∂ k on Y . To deal with non-diagonal insertions we express the invariance
of the integral Y under an infinitesimal change of variable of the form

M → M + iǫ[X, M ], with Xab = δak δbl Mkl . (4.4)

The Jacobian is 1 + iǫ(Mll − Mkk ), while the term trM 3 is invariant. Thus
D i E
0 = Mll − Mkk + (λ2k − λ2l )Mkl Mlk (4.5)
2
29
with no summation implied. Inserting this into (4.3) leads to3

X hMkk − Mll i
0 = λ2k + hMkk
2
i − 2i . (4.6)
λ2k − λ2l
l,l6=k

This yields the matrix Airy equations

h  2 X  i
2 1 ∂ 1 1 ∂ 1 ∂
λk − −2 − Y =0 (4.7)
λk ∂λk λ2k − λ2l λk ∂λk λl ∂λl
l,l6=k

which can be turned into equivalent equations for Z itself

 !2
X 1 X  
 1 1
+ 4 +2
1 1 ∂
−
1 ∂
λ2k λl 4λk λ2k − λ2l λk ∂λk λl ∂λl
l l,l6=k
!  2 #
1 X 1 ∂ 1 ∂
−2 1 + 2 + Z=0. (4.8)
λk λk + λl ∂λk λk ∂λk
l

In the limit N → ∞ we know from sec. 3 that Z admits an expansion in terms of odd
traces
X 1
tn = −(2n − 1)!! . (4.9)
l
λl2n+1

The differential equations (4.7) can be expanded in inverse powers of λk in the form
X 1
2 Lm Z = 0. (4.10)
(λ2k )m+2
m≥−1

Explicitly

1 X ∂ ∂
L−1 = t20 + tk+1 −
2 ∂tk ∂t0
k≥0

3
For an alternative derivation one can transform the “equation of motion” (4.3) into a matrix
differential equation, assuming at first the argument T ≡ Λ2 to be an arbitrary (i.e. not necessarily
diagonal) Hermitian matrix. Recognizing that the integral is invariant under conjugation of T ,
hence only a function of its eigenvalues {ta } one then uses

∂ X ∂ta ∂ X Minkl (ta − T ) ∂

= =
∂Tkl ∂Tkl ∂ta P ′ (ta ) ∂ta
a a

where P (x) = det(x − T ) and Minkl denotes the (k, l) minor in the corresponding matrix.

30
 1 X ∂ ∂
L0 = + (2k + 1)tk −3 (4.11)
8 ∂tk ∂t1
k≥0
X ∂ 1 ∂2 ∂
L1 = (2k + 1)(2k − 1)tk−1 + 2 − 15
∂tk 2 ∂t0 ∂t2
k≥1
X ∂ ∂2 ∂
L2 = (2k + 1)(2k − 1)(2k − 3)tk−2 +3 − 105
∂tk ∂t0 ∂t1 ∂t3
k≥2
.........
X (2k + 1)!! ∂ 1 X ∂2
Lm = tk−m + (2k + 1)!!(2l + 1)!!
(2(k − m) − 1)!! ∂tk 2 ∂tk ∂tl
k≥m k+l=m−1

∂ t20 1
− (2m + 3)!! + δm+1,0 + δm,0
∂tm+1 2 8

In eq. (4.10), each coefficient has to vanish so that

Theorem 2 (Kontsevich)
Z satisfies and is determined by the highest weight conditions

Lm Z = 0 m ≥ −1. (4.12)

The operators Lm obey (part of) the Virasoro (or rather the Witt) algebra, namely

[Lm , Ln ] = (m − n)Lm+n m, n ≥ −1 (4.13)

generated by L−1 , L0 , L1 , L2 . Note that only the first two involve first order derivatives.

5. Genus expansion.

As in standard matrix models there exists a genus expansion for F

X
ln Z = F = Fg (5.1)
g≥0

One way to obtain it is from the Airy system. One inserts appropriate extra factors of
N and studies the large N limit, paying attention to corrections, according to a method
applied to leading order by Kazakov and Kostov [6] and revived by Makeenko and Semenoff
[5]. In the spirit of our paper we follow a slightly different approach based on the KdV

31
equations and Virasoro constraints. In genus g, Fg collects in the expansion (2.1) all terms
such that * +
X Y (τi ti )ki
Fg (t. ) = (5.2)
ki !
ki i≥0
Σi≥0 (i−1)ki =3g−3

We set
∂ 2 Fg
ug (t. ) = . (5.3)
∂t20
In the KdV hierarchy (eq. (2.9)) the leading term in the semi-classical or genus expansion
corresponding to the term u0 is obtained from power counting by ignoring in the differential
polynomial Rn all terms involving derivatives. It then reduces to

∂u0 ∂ un+10
n≥0 = . (5.4)
∂tn ∂t0 (n + 1)!

For n = 0 this is vacuous and has to be supplemented by the first Virasoro condition (4.10)
(for m = −1) which amounts to

∂u0 X ∂u0
=1+ tk+1 . (5.5)
∂t0 ∂tk
k≥0

Inserting (5.4) into (5.5) gives

 
∂  X u n
u0 − tn 0 = 0 . (5.6)
∂t0  n! 
n≥0

Define
X up0
Ik (u0 , t. ) = tk+p . (5.7)
p!
p≥0

Equation (5.6) suggests

Lemma 4
u0 (t. ) satisfies the implicit equation

u0 − I0 (u0 , t. ) = 0. (5.8)

un
To check this, multiply (5.5) by 0
n! and use (5.4) to obtain

∂u0 un0 X ∂ un+10

= + tk+1 (5.9)
∂tn n! ∂tk (n + 1)!
k≥0

32
while
∂I0 (u0 , t. ) un X ∂ uk+1
= 0 + tk+1 0
. (5.10)
∂tn n! ∂tn (k + 1)!
k≥0

Subtracting and using (5.4) again, we find

 
∂ h i X un0 ∂ uk ∂
u0 − I0 (u0 , t. ) = tk+1 u0 − 0 u0 (5.11)
∂tn n! ∂tk k! ∂tn
k≥0
!
X un0 ∂ uk+1 uk0 ∂ un+1
0 0
= tk+1 − =0.
n! ∂t0 (k + 1)! k! ∂t0 (n + 1)!
k≥0

The difference u0 − I0 (u0 , t. ) is thus a constant. Since it vanishes at t. = 0 it is identically

zero, completing the proof. To find F0 we have to integrate twice eq. (5.8) with respect to
∂
t0 . From (5.2) the boundary conditions are given by the vanishing of F0 and ∂t0 F0 when
t0 = 0. This yields with u0 (t. ) implicitly given by (5.8)

u30 X uk+2
0 tk 1 X uk+1
0
X ta tb
F0 = − + . (5.12)
6 k + 2 k! 2 k+1 a! b!
k≥0 k≥0 a+b=k

Remarks
(i) As is generally the case if we extend F0 by considering u0 (originally equal to ∂ 2 F0 /∂t20 )
as an auxiliary independent parameter, we find that the stationarity condition

∂F0 1 2
= (u0 − I0 (u0 , t. )) = 0 (5.13)
∂u0 2

yields equation (5.8).

(ii) The expression for F0 is equivalent to the one given by Makeenko and Semenoff [5]
using (infinitely many) eigenvalues λk

1X 3 1X 3
X 1
F0 = λk − (λk − 2s) 2 − s (λ2k − 2s) 2
3 3
k k k
p p
s 3
1 X λk − 2s + λ2l − 2s
2
+ − ln (5.14)
6 2 λk + λl
k,l

with the condition !

∂F0 1 1 X
= s+ p =0 (5.15)
∂s 2 λ2 − 2s
k
k
P
upon identification of s with u0 , of Ip with −(2p − 1)!! k 2 1 1 and tn is as in (4.9).
(λk −2s)p+ 2

33
(iii) From (5.12) or (5.8) we can readily find the first few terms in the expansion of F0
which up to a factor t30 only involves the combinations tk−1
0 tk .
 4   5 
t30 t30 t0 t21 t30 t0 t40 t30 t31
F0 = + t1 + t2 + 2 + t3 + 3t1 t2 + 6
3! 3! 4! 2! 3! 5! 4! 3! 3!
 6  2  5 3 4 2 4

t t t t t t t
+ t4 0 + 6 2 + 4t1 t3 0 + 24 0 1 + 12t2 1 0
6! 2! 5! 3! 4! 2! 4!
 7 6 3 5
  
t0 t0 t0 t1 t22 t21 t50 t31 t40
+ t5 + (5t1 t4 + 10t2 t3 ) + 120 + 30t1 + 20t3 + 60t2
7! 6! 3! 5! 2! 2! 5! 3! 4!
 8  2  7 3 6
 3 2
 6
t0 t3 t0 t0 t1 t2 t1 t
+ t6 + 20 + 6t1 t5 + 15t2 t4 + 720 + 90 + 30t4 + 60t1 t2 t3 0
8! 2! 7! 3! 6! 3! 2! 6!
 3 2 2
 5 4 4

t t t t0 t t
+ 120t3 1 + 180 1 2 + 360t2 0 1 + . . . (5.16)
3! 2! 2! 5! 4! 4!

In genus zero all coefficients are positive integers (as opposed to fractional) due to the
smoother structure of M0,n , (n ≥ 3). Indeed we have the obvious
Lemma 5
The class of formal power series in t0 , t1 , . . . which vanish at t. = 0 with non-negative
integral derivatives at the origin is stable under
(i) addition
(ii) product
(iii) composition
To apply this to u0 (and hence to F0 ) we remark that the sequence
∞
X
k tk
f 0 = t0 , fn = fn−k (5.17)
0
k!

has each of its derivatives at the origin which stabilizes to the corresponding one of u0
after finitely many steps.
To obtain the next terms we split the Virasoro constraints expressed on ln Z as follows

m = −1
t20 X ∂Fg ∂Fg
δg,o + tk+1 − =0
2 ∂tk ∂t0
k≥0
m=0
1 X ∂Fg ∂Fg
δg,1 + (2k + 1)tk −3 =0 (5.18)
8 ∂tk ∂t1
k≥0
m=1

34
 X ∂Fg ∂Fg 1 ∂ 2 Fg−1 1 X ∂Fg1 ∂Fg2
(2k + 1)(2k − 1)tk−1 − 15 + + =0
∂tk ∂t2 2 ∂t20 2 g +g =g ∂t0 ∂t0
k≥1 1 2

m=2
X ∂Fg ∂Fg ∂ 2 Fg−1 X ∂Fg ∂Fg
1 2
(2k + 1)(2k − 1)(2k − 3)tk−2 − 105 +3 +3 =0,
∂tk ∂t3 ∂t0 ∂t1 g +g =g
∂t 0 ∂t 1
k≥2 1 2

while a similar splitting of the KdV equation (2.9) yields

!
∂ug ∂ 1 ∂ 2 ug−1 1 X
= + ug ug . (5.19)
∂t1 ∂t0 12 ∂t20 2 g +g =g 1 2
1 2

For genus one this equation reads

  
∂ ∂ ∂ ∂F1 1 ∂ 2 u0
− u0 − = 0. (5.20)
∂t0 ∂t1 ∂t0 ∂t0 12 ∂t20

We have

∂u0 1 ∂u0 u0
= , =
∂t0 1 − I1 ∂t 1 − I1
1 
∂Ip Ip+1 ∂ ∂
p≥1 = , − u0 Ip = δp,1
∂t0 1 − I1 ∂t1 ∂t0

hence we can rewrite

1 ∂ 2 u0 1 I2 1 ∂ I2
2 = 3
=
12 ∂t0 12 (1 − I1 ) 24 ∂I1 (1 − I1 )2
 
∂ ∂ 1 1
= ln . (5.22)
∂I1 ∂t0 24 1 − I1

If F1 is a function of t. only through I1 , I2 , . . ., we can rewrite

 
∂ ∂ ∂ ∂ ∂
− u0 F1 = F1 (5.23)
∂t1 ∂t0 ∂t0 ∂I1 ∂t0

so that equation (5.20) becomes

 
∂ ∂ ∂ 1 1
F1 − ln = 0. (5.24)
∂t0 ∂I1 ∂t0 24 1 − I1

This suggests that

1 1
F1 = ln , (5.25)
24 1 − I1

35
in agreement with the above hypothesis so that eq. (5.19) is satisfied. A straightforward
computation shows that the Virasoro conditions are satisfied, proving (5.25).
Remark
It is not unexpected that the genus one (or “one-loop”) result involves as usual a logarithm.
Expanding F1
   3 
t21 t1 t20
24F1 = t1 + + t0 t2 + 2 + t3 + 2t0 t1 t2
2! 3! 2!
 4 
t1 t30 t20 t22 t21 t20
+ 6 + t4 + 4 + 6t0 t2 + 3t1 t3
4! 3! 2! 2! 2! 2!
 5 4 3

t1 t0 t1 t30 t20 t22 t20 t21
+ 24 + t5 + 24t0 t2 + (4t1 t4 + 7t2 t3 ) + 16t1 + 12t3
5! 4! 3! 3! 2! 2! 2! 2!
  
t6 t5 t4 t2 t4 t3 t3 t2 t3
+ 120 1 + t6 0 + 120t0 t2 1 + 14 3 + 5t1 t5 + 11t2 t4 0 + 48 0 2 + 60t3 0 1
6! 5! 4! 2! 4! 3! 3! 2! 3!
 2
 3 2 2 2

t t t t t
+ 20t4 1 + 35t1 t2 t3 0 + 80 0 1 2
2! 3! 2! 2! 2!
 7 6 5
t1 t0 t1 t50 t20 t41
+ 720 + t7 + 720t0 t2 + (6t1 t6 + 16t2 t5 + 25t3 t4 ) + 360t3
7! 6! 5! 5! 2! 4!
 2 2 2
 4
t t t t
+ 84t1 3 + 118t3 2 + 30t5 1 + 66t1 t2 t4 0
2! 2! 2! 4!
3 3
 3 2 2
 3 
t0 t2 t0 t0 t2 t1 t21 t30
+288t1 + 120t4 + 480 + 210t2 t3 +... (5.26)
3! 3! 3! 2! 2! 3! 2! 3!
1
All intersection numbers are of the form 24
× (a positive integer) since I1 and − ln(1 − I1 )
belong to the class of functions referred to in Lemma 5.
For higher genus the Ansatz
l
X l3g−2 1 I2l2 I3l3
3g−2
I3g−2
Fg = hτ2l2 τ3l3 . . . τ3g−2 i P ... , (5.27)
P (1 − I1 )2(g−1)+ lp l2 ! l3 ! l3g−2 !
(k−1)lk =3g−3
2≤k≤3g−2

2k+1
which is a finite sum of monomials in Ik /(1 − I1 ) 3 , the number of which is p(3g − 3)
(with p(n) the number of partitions of n), is consistent with the KdV equation (5.19).
Inserted into the latter, it allows one to compute the coefficients with the result
 
1 I4 I3 I2 I23
F2 = 5 + 29 + 28 . (5.28)
5760 (1 − I1 )3 (1 − I1 )4 (1 − I1 )5
Hence
1 29 7
hτ4 i = hτ2 τ3 i = hτ23 i = (5.29)
1152 5760 240
36
in agreement with Witten [1]. The other intersection numbers can be derived by expanding
(5.28) in t. .
For genus 3 we find

1 I7 I6 I2 I5 I3 I5 I22
F3 = 35 + 539 + 1006 + 4284
2903040 (1 − I1 )5 (1 − I1 )6 (1 − I1 )6 (1 − I1 )7
I42 I4 I3 I2 I4 I23 I33
+607 + 13452 + 22260 + 2915
(1 − I1 )6 (1 − I1 )7 (1 − I1 )8 (1 − I1 )7

I32 I22 I3 I24 I26
+43050 + 81060 + 34300 , (5.30)
(1 − I1 )8 (1 − I1 )9 (1 − I1 )10
which yields the table

1 1121
hτ7 i = 82944 hτ4 τ3 τ2 i = 241920
77 53
hτ6 τ2 i = 414720 hτ4 τ23 i = 1152
503 583
hτ5 τ3 i = 1451520
hτ33 i = 96768
17 205
hτ5 τ22 i = 5760 hτ32 τ22 i = 3456
607 193
hτ42 i = 1451520 hτ3 τ24 i = 288
1225
hτ26 i = 144

Table I

There is no difficulty to pursue these computations as far as one wishes.

Remarks
(i) It follows from Lemma 5 that all series coefficients in (5.27) reexpressed in terms of
Q lk
tk /lk ! are non-negative integers up to the finitely many prefactors. All intersection
numbers of a fixed genus when written as irreducible fractions have therefore a lowest
common multiplier (l.c.m.) Dg : D0 = 1, D1 = 24, and for g > 1, Dg is the lowest
common denominator of the finitely many intersection numbers appearing in (5.27), when
written as irreducible fractions. We conjecture that for 1 < g ′ ≤ g, the order of any
automorphism group of an algebraic curve of genus g ′ (bounded by 84(g − 1)) divides Dg .
Thus D3 = 2903040 is divible by 168 (the order of the largest automorphism group of a
genus 3 curve) and by 48 (the same for genus 2).
(ii) The term in Fg , g > 1 which has the highest power of (1 − I1 ) in the denominator has
the form
hτ23g−3 i I23g−3
(3g − 3)! (1 − I1 )5g−5

37
In the next section we develop a formalism to resum these terms as well as subleading
ones akin to the “double scaling limit” of standard matrix models [10] (one recognizes the
same string exponents and the same ingredients). At the other extreme the term with the
lowest power of 1 − I1 in the denominator is [1]

I3g−2 1
hτ3g−2 i hτ3g−2 i = . (5.31)
(1 − I1 )2g−1 (24)g g!

The last equality (also valid for g = 1) follows from (5.19) by keeping terms with the lowest
power of (1 − I1 )−1 . It implies that (24)g g! divides Dg .

6. Singular behaviour and Painlevé equation.

The expression (2.18) exhibits a singular behaviour as I1 → 0. In a first step, we can keep
in the KdV equation the dominant terms by considering that

I2 ≈ constant, Ik ≈ 0 for k ≥ 3. (6.1)

This is consistent with the derivatives of the I’s: ∂Ik /∂t0 = Ik+1 /(1 − I1 ) and ∂Ik /∂t1 =
u0 Ik+1 /(1 − I1 ) for k ≥ 2. In this approximation, the genus g contribution to the specific
heat ug = ∂ 2 Fg /∂t20 is of the form

3(g−1)+2
2 I2
ug = ∂ Fg /∂t20 = αg . (6.2)
(1 − I1 )5(g−1)+4

We introduce the scaling variable

(1 − I1 )
z= 3/5
(6.3)
I2
P
e, u
and separate the genus zero contribution by setting u = u0 + u e= g≥1 ug . The KdV
equation (2.9a) then reads

∂e
u 1 u
1
∂e I22 1 ∂ 3u
e
=e
u + + + 3
∂t1 1 − I1 ∂t0 4 (1 − I1 )5 12 ∂t0
1 u
1
∂e I22 1 u
3
I2 ∂e
=e
u + + + . (6.4)
1 − I1 ∂I1 4 (1 − I1 )5 12 1 − I1 ∂I1

A rescaling
−2/5
e = I2
u ψ(z) (6.5)

38
leads to the equation

∂ψ ψ  ∂ψ  1 1 1 ∂
3
+ 1− + 5− ψ = 0. (6.6)
∂z z ∂z 4z 12 z ∂z

This equation will be generalized below. In this particular case of the behaviour (6.1), one
can transform it into the Painlevé equation: we set t = 2−2/5 z 2 , ψ(z) = z + 21/5 φ(t) and
find
1 ′′
φ + φ2 − t = 0 . (6.7)
3
which has the asymptotic expansion

X φg
φ= 5 , φ0 = −1 , (6.8)
t 2 (g−1)+2

where the successive terms satisfy

g
25g 2 − 1 1 X
φg+1 = φg + φg+1−m φm . (6.9)
24 2 m=1

Hence in this regime

X X hτ 3g−3 i 3(g−1)
I2
Fgsing = 2
(3g − 3)! (1 − I1 )5(g−1)
g≥2 g≥2
2g (3g − 3)!
hτ23g−3 i = φg (6.10)
(5g − 5)(5g − 3)

7 1225 1816871
g = 2 hτ23 i = , g = 3 hτ26 i = , g = 4 hτ29 i = ,... . (6.11)
240 144 48

This discussion may be extended to the regime in which all (or a finite number of)
the I’s are retained and tend to zero according to the following scaling law

(1 − I1 )
z= 3/5
I2
Iq (1 − I1 )q−2
vq = q≥3 (6.12)
I2q−1
3g−2
X l3g−2
Y vqlq
Fgsing =z −5(g−1)
hτ2l2 τ3l3 . . . τ3g−2 i
l !
q=3 q
Σ2≤k≤3g−2 (k−1)lk =3g−3

39
The KdV equation is then rephrased as

h ∂ X ∂ i
z + (q − 2)vq +1 ψ = (6.13)
∂z ∂vq
q≥3
1 2
3 + v3 1 8 3 2
= ψ ∆ + v3 ψ − 4
+ 5
[∆ − 4 − v3 ][∆ − 2 − v3 ][∆ + v3 ]ψ
z 5 12z 12z 5 5 5

where the same change of function as in (6.5) has been carried out, and ∆ denotes the
differential operator

3 ∂ X
∂
∆ = (1 + v3 )z + − vq+1 + (q − 2)vq + (q − 1)vq v3 . (6.14)
5 ∂z ∂vq
q≥3

The contribution to a given genus g involves only a finite number of terms in the sum:
q ≤ 3g − 2. Moreover the differential operators in (6.13) respect the grading in powers of
z,

ψg (z, v. ) = z −5(g−1)−4 ψg (v. ) ,

thus determining the polynomials ψg (v. ) recursively. For instance, if only z and v ≡ v3
are retained, we have for g > 1 (and ∂v ≡ ∂/∂v)

g−1
X
[5(g − 1) + 3 − v∂v ] ψg (v) = ψg−g ′ (v) [4 + 2v + (g ′ − 1)(5 + 3v) − (1 + 2v)v∂v ] ψg ′ (v)
g ′ =1
1
+ [8 + 4v + (g − 2)(5 + 3v) − (1 + 2v)v∂v ] × (6.15)
12
[6 + 3v + (g − 2)(5 + 3v)− (1 + 2v)v∂v ] [4 + 2v + (g − 2)(5 + 3v) − (1 + 2v)v∂v ] ψg−1 (v) ,

which yields

24 ψ1 (v) =2 + v

1152 ψ2 (v) =196 + 352v + 109v 2 (6.16)

82944 ψ3 (v) =117600 + 362564v + 324660v 2 + 84699v 3 + 3043v 4

3981312 ψ4 (v) =1906157232 + 7865959024v + 11212604992v 2 + 6581090736v 3

+ 1465796801v 4 + 83580341v 5 .

40
From this one extracts the first intersection numbers of the form hτ2l2 τ3l3 i. One recovers
for g ≤ 3 results obtained in (5.29) or in Table I, and in genus g = 4 the results

1816871
hτ29 i =
48
3326267
hτ27 τ3 i =
1728
5 2 728465
hτ2 τ3 i =
6912
43201
hτ23 τ33 i =
6912
134233
hτ27 τ34 i = .
331776

Table II
We hope we have amply demonstrated the practical use of these expansions.

7. Generalization to higher degree potentials

The cubic potential in the integral (0.1) may be generalized to a potential of degree p + 1,
as noticed by several authors [7],[2],[8]. The case p = 2 is the one discussed previously.
Let us consider first the one-variable integral analogous to (2.12), also denoted z(λ). The
normalizations are adjusted in such a way as to make the quadratic term positive definite,
in order to have a well-defined asymptotic expansion
2  
R∞ ip +1
(m+(−i)p+1 λ))
p+1

−∞
dm e 2(p+1) >lin
z(λ) = R∞ p 2 λp−1 , (7.1)
−∞
dm e− 4 m

where the subscript “> lin” denotes the sum of terms of degree ≥ 2 in the polynomial. By
considerations similar to those of sect. 2.1, it is easy to see that z(λ) admits an asymptotic
expansion in inverse powers of λp+1
∞
X
z(λ) = ck λ−(p+1)k (7.2)
0

with c0 = 1. It satisfies a differential equation of order p + 1

(Dp − λp )z(λ) = 0 (7.3)

41
  
p+1 ∂
p p
p−1
− 2(p+1) (−λ)p+1 − p−1 (−λ)p+1
D =λ 2e 2(−1) λ 2 e 2(p+1)
∂λp
 
(−1)p p − 1 2 ∂
=λ + p
− p−1 . (7.4)
p λ λ ∂λ

The corresponding matrix integral

2  
R ip +1
tr (M +(−i)p+1 Λ))
p+1

dM e 2(p+1) >lin
Z (N) (Λ) = R 1
Pp−1 k p−1−k
(7.5)
dM e− 4 tr k=0
MΛ MΛ

may then be handled as in equations (2.17)–(2.24). One considers the set of functions z (j)
defined by
z (j) (λ) = λ−j Dj z(λ), j≥0. (7.6)

From eq. (7.3), it follows that

Dr(p−1)+j z = λr(p−1)+j z (j) mod (D0 z, . . . , Dr(p−1)+j−1 z).

Thus the analogue of (2.24) reads

|λ0 z (0) , . . . , λp−1 z (p−1) , λp z (0) , . . . |

Z (N) (Λ) = . (7.7)
|λ0 , λ1 , λ2 , · · · , λN−1 |

The p functions z = z (0) , . . . , z (p−1) have asymptotic expansions

∞
X (j)
z (j) (λ) = ck λ−k(p+1) , (7.8)
0

(j)
with coefficients ck ; in the sequel the latter are regarded as periodic in j of period p:
(j+p) (j)
ck = ck . One then proceeds as in sec. 2.2, introducing Schur functions, with the
result (analogous to (2.41)) that

p(p+1)n0
X
(N)
Zk (Λ) = c(0) (1) (N−1) ..
n0 cn1 · · · cnN −1 . (7.9)
n0 +···+nN −1 =k
p(p+1)nN −1

is independent of N for N ≥ (p + 1)k.

The steps of sec. 2.2 may then be followed sequentially to prove that Zk , computed
now for N = (p + 1)k, is independent of the θrp = trΛ−rp . When differentiating Zk with
respect to θrp , the only non-trivially non vanishing terms are those for which the derivative

42
acts on one of the last rp lines, where r is a multiple of p + 1. The discussion of such a
case then appeals to the following identity generalizing (2.56)
h i
ij (j) i
det ω z (ω λ) = constant (7.10)
0≤i,j≤p−1

where ω is a p-th root of unity. This is proved by differentiating the determinant, using
the relations (7.6) between the functions z (j) . This implies a family of identities on the
coefficients
X
Ckp ≡ ǫπ c(0) (p−1)
n0 . . . cnp−1 = 0 (7.11)
P
ni =kp
ni =i+π(i) mod p

where the summation runs over the configurations of indices ni that may be written as
indicated, with π a permutation of the p integers 0, . . . , p − 1.
∂Z
On the other hand, as before, the only contributions to ∂θrp(p+1) come from Young
tableaux with a square-rule shape (as in eq. (2.60)), and one finds that
 
rp−1
X rp−1−l
X p−1
X
∂Z 2c(i) ∆(i+j+1) + (i+j) (i+j+1) 
= (−1)l rp−l l crp−l ∆l , (7.12)
∂θrp(p+1)
l=0 i=0 j=1

where the ∆’s are determinants generalizing those of (2.63):

(j) (j+1) (j+s−1)
c1 c2 . . . cs
(j+1) (j+s−1)
1 c1 . . . cs−1
∆s(j) = . .. .. . (7.13)
.. . .
(j+s−1)
0 ... 1 c1
The expression (7.12) may be recast in the form

∂Z X p−1
r X h p−1−i
X
Xp−1 i
(s−1)p+i (j) (j+1)
= (−1) (r−s+1)(p+1)−i + c(r−s+1)p−i ∆i+(s−1)p .
∂θrp(p+1) s=1 i=0 j=0 j=p−i
(7.14)
(j) (j+1)
It appears that the combination of c and ∆ in the summand in (7.14), namely csp−i ∆i ,
(j)
is, up to a sign, the coefficient of csp−i in the constraint Csp of eq. (7.11)
(j) (j+1) (p−1)(p−2) X′
+i
csp−i ∆i = (−1) 2 ǫπ c(0) (p−1)
n0 . . . cnp−1 (7.15)
P′
with the sum subject to the same constraints as in (7.11) and to nj = sp − i. Using
this fact and after some reshuffling, one finds that ∂Z/∂θrp(p+1) is proportional to the
constraint Crp and thus vanishes,
∂Z (p−1)(p−2)
= (−1) 2 r(p + 1)Crp = 0. (7.16)
∂θrp(p+1)

43
The last part of the argument is carried out as in the end of sec. 2.2, thus completing the
proof of the independence of Z with respect to the θrp .
One then proceeds as in sec. 3, deriving the higher KdV hierarchies associated with a
differential operator Qp of order p depending on p − 1 functions, and as in sec. 4, writing
the generalized Airy equation satisfied by Z. For example, in the case p = 3, we have
n 1 X  1 1

3 2 2
− tj + D j + Dj (Dj − Dk ) + (D − Dk )
8 tj − tk tj − tk j
k,k6=j
X  1
o
+ Dj + circ. Z=0 (7.17)
(tj − tk )(tj − tl )
k,l
j6=k6=l6=k

with the following notations

tj = λ3j
∂
∂j =
∂tj
3 4 Y
12 3 4 Y
− 21
Dj = e− 8 Σk λk λ2k + λk λl + λ2l ∂j e 8 Σk λk λ2k + λk λl + λ2l
k,l k,l
= ∂j + a j (7.18)
1 1 X 2λj + λk
aj = λj − 2
2 3λj λj + λk λj + λ2k
2
k

From this system of equations, the strong reader will be able to extract the expression of
the generators of the W3 algebra, in a way similar to (4.10), and to calculate the analogues
of the genus expansion and of the singular behavior discussed in sec. 5 and 6 . . .

Acknowledgements.
It is a pleasure to acknowledge some inspiring correspondence with M. Kontsevich and to
thank M. Bauer for his assistance in algebraic calculations as well as in the elaboration of
Lemma 5 and P. Ginsparg for a critical reading of the manuscript.

44
 References

[1] E. Witten, Two dimensional gravity and intersection theory on moduli space, Surv. in
Diff. Geom. 1 (1991) 243-310.
[2] M. Kontsevich, Intersection theory on the moduli space of curves and the matrix Airy
function, Bonn preprint MPI/91-77.
[3] E. Witten, On the Kontsevich model and other models of two dimensional gravity,
preprint IASSNS-HEP-91/24
[4] E. Witten, The N matrix model and gauged WZW models, preprint IASSNS-HEP-
91/26, to appear in Nucl. Phys. B.
[5] Yu. Makeenko and G. Semenoff, Properties of hermitean matrix model in an external
field, Mod. Phys. Lett. A6 (1991) 3455-3466.
[6] V. Kazakov and I. Kostov, unpublished ;
I. Kostov, Random surfaces, solvable lattice models and discrete quantum gravity in
two dimensions, Nucl. Phys. B (Proc. Suppl.) 10A (1989) 295-322.
[7] M. Adler and P. van Moerbeke, The Wp –gravity version of the Witten–Kontsevich
model, Brandeis preprint, September 1991
[8] E. Witten, Algebraic geometry associated with matrix models of two dimensional grav-
ity, preprint IASSNS-HEP-91/74.
[9] V. Kazakov, The appearance of matter fields from quantum fluctuations of 2D-gravity,
Mod. Phys. Lett. 4A (1989) 2125-2139.
[10] E. Brézin and V. Kazakov, Exactly solvable field theories of closed strings, Phys. Lett.
236 (1990) 144-150;
M. Douglas and S. Shenker, Strings in less than one dimension, Nucl. Phys. B335
(1990) 635-654;
D.J. Gross and A. Migdal, Non perturbative two–dimensional quantum gravity, Phys.
Rev. Lett. 64 (1990) 127-130.
[11] C. Itzykson and J.-B. Zuber, The planar approximation II, J. Math. Phys. 21 (1980)
411-421.
[12] Harish-Chandra, Differential operators on a semisimple Lie algebra, Amer. J. Math.
79 (1957) 87-120.
[13] F.D. Murnaghan, The Theory of Group Representations, The Johns Hopkins Press,
Baltimore, 1938 ;
H. Weyl, The Classical Groups, Princeton Univ. Press, Princeton, N.J., 1946.
[14] M. Sato, Soliton equations as dynamical systems on an infinite dimensional Grass-
mann manifold, RIMS Kokyuroku, 439 (1981) 30-46 ;
The KP hierarchy and infinite-dimensional Grassmann manifolds,Proc. Symp. Pure
Math. 49 (1989) 51-66;

45
 M. Sato and Y. Sato, Soliton equations as dynamical systems on an infinite dimen-
sional Grassmann manifold, in Non linear PDE in Applied Science, US-Japan Semi-
nar, Tokyo 1982, Lecture Notes in Num. Appl. Anal. 5 (1982) 259-271.
[15] V.G. Drinfeld and V.V. Sokolov, Lie algebras and equations of the Korteweg–de Vries
type , Journ. Sov. Math. 30 (1985) 1975-2036.

46