Random orthonormal bases of spaces of high dimension

RANDOM ORTHONORMAL BASES OF SPACES OF HIGH DIMENSION

STEVE ZELDITCH

Abstract. We consider a sequence T-L n of finite dimensional Hilbert spaces of dimensions
djv — ► oo. Motivating examples are eigenspaces, or spaces of quasi- modes, for a Laplace or
Schrodinger operator on a compact Riemannian manifold. The set of Hermitian orthonormal
bases of "Hat may be identified with [/(d/v), and a random orthonormal basis of ® w Hn is
a choice of a random sequence Un € U(d^) from the product of normalized Haar measures.
We prove that if — > oo and if j-Tr A\-u N tends to a unique limit state uj(A), then almost
surely an orthonormal basis is quantum ergodic with limit state w(A). This generalizes an
earlier result of the author in the case where is the space of spherical harmonics on S 2 .
In particular, it holds on the fiat torus M. d /Z d if d > 5 and shows that a highly localized
orthonormal basis can be synthesized from quantum ergodic ones and vice- versa in relatively
small dimensions.

The purpose of this article is to prove a general result on the quantum ergodicity of random
orthonormal bases {ipN,j}j=i of finite dimensional Hilbert spaces "Hat C L 2 (M) of dimensions
d,N — > oo of a compact Riemannian manifold (M, g). The proof is based on a "moment poly-
tope" interpretation of quantum ergodicity from |Zlj : the quantum variances of a Hermitian
observable A e \1/°(M) are identified with moments of inertia of the convex polytopes V?

defined as the convex hull of the vectors A = (Ai, . . . , A^) of eigenvalues (in all possible
orders) of UnAU^ where : L 2 (M) — > %n is the orthogonal projection. Equivalently,
is the image of the coadjoint orbit of the diagonal matrix -D(A) under the moment map
for the Hamiltonian action of the maximal torus Tj N C U(cIn) of diagonal matrices acting
by conjugation on In particular, the main estimates of quantum ergodicity can be for-
mulated in terms of estimates of the first four moments of inertia of Vt. The main result,
Theorem [TJ states that random orthonormal bases are almost surely quantum ergodic as
long as (In — » oo and ^TrU N AU N — > oj(A) for all A e \I/°(M), where lo(A) is the Liouville
state. More generally, if these traces have any unique limit state, then almost surely it is the
quantum limit of a random orthonormal basis. The proof is essentially implicit in [Zlj, but
we bring it out explicitly here and also give detailed calculations of the moments of intertia,
which seem of independent interest.

Quantum ergodicity of random orthonormal bases is a rigorous result on the 'random
wave model' in quantum chaos, according to which eigenfunctions of quantum chaotic sys-
tems should behave like random waves. It also has implications for the approximation of
modes by quasi-modes. Since eigenfunctions of the Laplacian A of a compact Riemannian
manifold (M , g) form an orthonormal basis, it is natural to compare the orthonormal basis of
eigenfunctions to a 'random orthonormal basis'. In |Z1] . the result of this article was proved
for the special case where T-Ln is the space of degree N spherical harmonics on the standard

Research partially supported by NSF grants # DMS-0904252 and DMS-1206527.

2

STEVE ZELDITCH

S 2 . In |Z2| the quantum ergodic property was generalized to any compact Riemannian mani-
fold, with V,n the span of the eigenfunctions in a spectral interval [N, N+l] for vA- Related
results have recently been proved in |SZ[ IBL] . The dimension of such Hn grows at the rate
jV m_1 where m = dimM and thus a random element of Hn is a superposition of jV m-1
states. The results of this article show that the same quantum ergodicity property holds for
sequences of eigenspaces (or linear combinations) whose dimensions tend to infinity at
any rate. For instance, the results show that random orthonormal bases of eigenfunctions

on a flat torus of dimension > 5 are quantum ergodic (for the precise statement, see £4.1

and for further discussion, see £0.1

To explain the moment map interpretation and the variance formula, recall that quantum
ergodicity is concerned with quantum variances, i.e. with the dispersion from the mean of
the diagonal part of a Hermitian matrix Hn on a large dimensional vector space Hn- The
matrix Hn is the restriction

T$ ■= H N AU N (0.1)

to I-Ln of a pseudo-differential operator A 6 \I/ (M); here II/v is the orthogonal projection to
U N and *°(M) is the space of pseudo-differential operators of order zero. The same methods
and results apply to other context such as semi-classical pseudo-differential operators or to
Toeplitz operators on holomorphic sections of powers of a positive line bundle |SZj . Given an
ONB {ipN,j} d jZi of "Hat we define the quantum variances of the ONB (indexed by A e \E f0 (M)
by

VaMnh}) := -r yilWtfj,^) -co(A)\ 2 . (0.2)

dN U

Here, u>{A) = f s * M UAd^L where d\Xh is normalized Liouville measure (of mass one).
Definition: A sequence {ipNj}N of ONB's of %n is a quantum ergodic ONB of L 2 (M) if

(SV) lim V A ({ip Nk }) = 0, MA e *°(M). (0.3)

N-¥oo

By a standard diagonal argument, this implies that almost all the individual elements
(AipNjjipNj) tend to u)(A). Since this aspect of quantum ergodicity is the same as in

31 EZ] (e.g.) we do not discuss it here.
To define random orthonormal bases, we introduce the probability space (OAfB, dv), where
ONB is the infinite product of the sets OJ\fBjy of orthonormal bases of the spaces T-Ln, and
v = rijv=i u n, where un is Haar probability measure on ONBn- A point of ONB is thus a
sequence VP = {(ipi , . . . , V'^ v )}A r >i of orthonormal basis. Given one orthonormal basis {e^}
of "Hat any other is related to it by a unique unitary matrix. So the probability space is
equivalent to the product

oo

(ONB, dv) ~ Y[ [U{d N ), dU) (0.4)

N=l

where dU is the unit mass Haar measure on U(cIn)- Here we are working with Hermitian
orthonormal bases and Hermitian pseudo-differential operators. We could also work with real
self-adjoint operators and real orthornormal bases, which are then related by the orthogonal
group. The results in that setting are essentially the same but the proofs are somewhat more
complicated; for expository simplicity we stick to the unitary Hermitian framework.

RANDOM ORTHONORMAL BASES OF SPACES OF HIGH DIMENSION

3

Let A £ \J r ° and denote the eigenvalues of T- by Ai, . . . , X dN . The empirical measure of
eigenvalues of T- is defined by

(In

><r ! £*V (0-5)

A

N d N .

Its moments are given by

>N

Pk(Xi, ... } X dN ) = J2 X " = Tr(T^) fc . (0.6)
i=i

To obtain quantum ergodicity, we we put the following constraint on the sequence {Hn}'-
Definition: We say that Hn has local Weyl asymptotics if, for all A £ \I/ (M),

^TrT^ = u;(A)+o(l). (0.7)

In fact, the results generalize to the case where u)(A) is replaced by any other limit state,
i.e. f stM aAd(jL where d\i is another invariant probability measure for the geodesic flow.
Our main result is:

Theorem 1. Let be a sequence of subspaces of L 2 (M) of dimensions d^ = dim "Hat — >
oo. Assume that j^TtUnAUn = oj(A) + o(1) for all A £ \I/°(M). Then with probability one
in (ONB,dv), a random orthonormal basis of @ n 1-Ln is quantum ergodic.

A natural question (which we do not study here) is whether a random orthonormal basis
is QUE, i.e. whether

max{\ (Aip NJ ,ip NJ ) - uj(A)\ 2 , j = 1, . . . , d N } {a.s.)dv ?.

As a tail event the probability of a random orthonormal basis being QUE is either or 1.

We now explain how to formula Theorem [T] in terms of moment maps and polytopes.
Quantum ergodicity of a random orthonormal bases concerns the dispersion from the mean
of the diagonal part of The diagonal part depends on the choice of an orthonormal basis
of Hn- Once an orthonormal basis is fixed, iT$ can be identified with an element if at of
the Lie algebra u(d N ) of \J(d]y), and a unitary change of the orthonormal basis results in the
conjugation H N — > U n H n Un of H N . If the vector of eigenvalues of H N is denoted Ajv, then
the conjugates sweep out the orbit 0^ N - Let t(djv) denote the Cartan subalgebra of diagonal
elements in u(d N ), and let || ■ || 2 denote the Euclidean inner product on t(d N ). Also let

J dN : iu(d N ) -)■ it(d N )

denote the orthogonal projection (extracting the diagonal). Extracting the diagonal from
each element of the orbit is precisely the moment map

d

N

J dN :0 XN ^V XN Cit(d N ), J dN (UD(X)U*)=\^..,J2\ j \U ij \ 2 ,... ) (0.f

of the conjugation action of T dN . Finally, let

J dN (H) = Q-Tr HJ ld dN , D (X N ) = D(X N ) - Q-Tr H ) Id djv ,

4

STEVE ZELDITCH

for Hermitian matrices H 6 iu{d x ). We also introduce notation for the diagonal of D Q (\):

-D (A) = £>(A), with; A,- := A r -j-J] A,. (0.9)

Thus,

H = H [> + J d (H), resp. D(\ N ) = D (\ N ) + I J-Tr # ) Id (

A'

with if traceless, corresponds to the decomposition u(c? A r) = su{d x ) © E.

As this description indicates, quantum ergodicity of random orthonormal bases is mainly a
result about the asymptotic geometry of the polytopes V x corresponding to a sequence T$
of Toeplitz operators. The pushforward of the [/(ciAr)-invariant normalized measure on O x
to V x is the so-called Duistermaat-Heckman measure dC? H , a piecewise polynomial measure
on Vy To prove almost sure quantum ergodicity, we prove that for all such sequences T$
and their spectra {A^}, the second and fourth moments of inertia of V x with respect to

dCj? are bounded. We use the property in Definition 0.7 to replace oo(A) by the centers of

mass, i.e. the scalar matrix with the same trace as T-. The Kolmogorov strong law of large
numbers then gives the quantum ergodicity property. In [Z3J , we study higher moments and
their implication for the limit shape of V? along a sequence {Xn} with a limit empirical
measure.

We asymptotically evalute the moments using the Fourier transform

fa(X) := / e^ x ^ Y »d H (Y) (0.10)
Jo{\)

of the 5-function on Or. Here, we assume X 6 IR djv . We may identify X with a diagonal
matrix, and then (X, diag(F)) = TrXY, and we get the standard Fourier transform. We
obviously have:

Lemma 1. Let A be the Euclidean Laplacian ofR dN acting in the X variable. Then,

m 2 (V XN ) := E|| J dN (U*D(X)U)\\ 2 = -Afa(X)\ x=0 ,

m A {V lN ) := E|| J dN (U*D(\)U)\\* = A^ X (X)\ X=0 ,

We translate A by its center of mass to make the center of mass of V x equal to 0, i.e.
Aj = 0. Using a formula for (i\{X) in terms of Schur polynomials, we prove

Lemma 2. Letpk be the power functions (0.6). Assume that P\(X) = 0. Then,
A 2 /i x (0) = (3 4 (d N ) p!(A),

uri+h R (rl \ — ( ±d N {d N -l) ±d N {d N -l) , (12^+4^(^-1))

WU/l P4{UN) I (d N + l)d 2 N (d N -l) (d N +2)(d N +l)d N (d N -2) 1" (d N +3)(d N +2)(d N +l)d N

RANDOM ORTHONORMAL BASES OF SPACES OF HIGH DIMENSION

5

The proof of Theorem [I] follows directly from Lemma [2] and the Kolmogorov SLLN (strong
law of large numbers). When grows fast enough it also follows directly from the Borel-
Cantelli Lemma. We first introduce notation for the basic random variables:

Definition:

Y4:OATB N ^[0,+oo), V = (U dl ,U d2 ,...)

Y£(Hf) := \\J dN (U* N D(\)U N ) -D(\)f = || J djV (^L> (A)^

A'/

Then Lemma [T]-Lemma [2] determine the asymptotics of their mean and variance

Var (yrf) := E((Y^) 2 ) - (E((Y^)) 2 .

Corollary 1.

Var(Y^) = (f3 i (d N )- J ^)pl(A)^pi(A).

The Lemma first implies that E|| J dN (U*D(X)U)\\ 2 is bounded for all A G *°(M). Hence,
E(^-|| Jd JV (f/*-D(A)f/)|| 2 ) as long as —> oo. Thus, the mean of the quantum variances

(0.2[) tends to zero. As in [Zl, SZJ we then apply the Kolmogorov SLLN (or the martingale

convergence theorem). The {Y^} is a sequence of independent random variables as N varies
and Lemma [2] shows that they have bounded variance. Hence the SLLN implies that the
partial sums,

J2U Y n~ EY n) (0-11)

Sn

have the property,

„ N dr '

— Sn — > 0, almost surely (0.12)

and this is equivalent to quantum ergodicity of random orthonormal bases. As mentioned
above, if d^ grows at a faster rate one can obtain stronger results from the Borel-Cantelli
Lemma: E.g. if Y^=i i~ < 00 ' one obtains almost sure convergence j-Y^ — > (a.s.).

Since the argument above only requires that j-KY^ — > and Var(y^) is bounded, it
does not require any assumption that EY^ 4 tends to a limit. Our calculations therefore go
beyond what is necessary for almost sure quantum ergodicity, and pertain to the asymptotic
geometry of the polytopes V^. There is a natural condition on the this sequence of polytopes:

Definition: We say that the sequence {Hn} has Szego asymptotics if, for all A e \E' (M),
there exists a unique weak* limit, u% — > ua € .M(K) as iV — > oo. Here, .M(R) is the set of
probability measures on IR.

Under this stronger assumption, Lemma [2] gives moment asymptotics:
Proposition 1. Let Xn £ K djv be a sequence of vectors with the property that the empirical

measures (0.5) tend to a weak limit v. Then

™±CPx N )->4{L(t-i) 2 dvy

6

STEVE ZELDITCH

This Proposition is closely related to the "Weingarten theorem" that the matrix elements
y/dNUij are asymptotically complex normal random variables, where C/y are the matrix
elements of U G U{cLn) |Wj- Perhaps this explains why the fourth moment is a constant
multiple of the square of the second moment. It would be interesting to see if the pattern
continues; we plan to study V? further in [Z3J.

0.1. Discussion. The motivation for proving quantum ergodicity of random orthonormal
bases for Hn of any dimensions tending to infinity was prompted by the general question:
how many diffuse states (modes or quasi-modes) does it take to synthesize localized modes
or quasi-modes? Vice-versa, how many localized states does it take to synthesize diffuse
states? We would like to synthesize entire orthonormal bases rather than individual states
and measure the dimensions of the space of states in terms of the Planck constant h. Let us
consider some examples.

In the case of the standard S 2 , the eigenspaces Hn of A are the spaces of spherical harmon-
ics of degree N. They have the well-known highly localized basis of joint eigenfunctions
of A and of rotations around the X3-axis. By localized we mean that a sequence {Y^} with
m/N — > a microlocally concentrates on the invariant tori in S*S 2 where pe = a. Here,
Po(x,^) = where J| generates the x 3 -axis rotations. On the other hand, it is proved

in |Z1] that independent "random" orthonormal bases oH-Ln are quantum ergodic, i.e. are
highly diffuse in S*S 2 . Since dim?^ — 2iV + 1, it is perhaps not surprising that the same
eigenspace can have both highly localized and highly diffuse orthonormal bases when its
dimenson is so large. The question is, how large must it be for such incoherently related
bases to exist?

A setting where the eigenvalues have high multiplicity but of a lower order of magnitude
than on S 2 is that of flat rational tori R n /L such as R n /Z n . Of course it has an orthonormal
basis of localized eigenfunctions, e 1 ^'^ . The key feature of such rational tori is the high
multiplicity of eigenvalues of the Laplacian A of the flat metric. It is well-known and easy
to see that the multiplicity is the number of lattice points of the dual lattice L* lying
on the surface of a Euclidean sphere. We denote the distinct multiple A-eigenvalues by
Hn, the corresponding eigenspace hy %n and the multiplicity of n 2 N by cIn = dim "Hat. In
dimensions n > 5, (In ~ /^aT 2 , one degree lower than the maximum possible multiplicity of a
A-eigenvalue on any compact Riemannian manifold, achieved on the standard S n . Further,
j^TvUnAUn —> u)(A). Hence, the results of this article show that despite the relatively
slow growth of (In on a flat rational torus, orthonormal bases of "Ha? in dimensions > 5 are
almost surely quantum ergodic. The statement for dimensions 2,3,4 is more complicated
(see

An interesting setting where the behavior of eigenfunctions is largely unknown is that of
KAM systems. For these, one may construct a 'nearly' complete and orthonormal basis for
L 2 (M) by highly localized quasi-modes associated to the Cantor set of invariant tori. It
seems unlikely that the actual eigenfunctions are quantum ergodic; but the results of this
article show that if they resemble random combinations of the quasi-mode, then it is possible

that they are. Further discussion is in £4.2

RANDOM ORTHONORMAL BASES OF SPACES OF HIGH DIMENSION

7

I. Background

In this section, we review the definition of random orthonormal basis and relate it to
properties of the moment map for the diagonal action of the maximal torus Td N on co-
adjoint orbits of U(cIn).

1.1. Random orthonormal bases of eigenspaces. Suppose that we have a sequence of
Hilbert spaces Hs N = 1,2,... of dimensions d N = dim'Hjv — > oo. We define the large
Hilbert space

oo

' I X

N=l

and orthogonal projections

U N :U^U N . ri.r

We then consider the orthonormal bases (0.4 ) of H which arise from sequences of orthonormal
bases of %n .

1.2. The basic random variables. Let A e \1/°(M) be a zeroth order pseudo-differential
operator. By a Toeplitz operator we mean the compression T- (0.1) of A to %n-

P A
■ N

Given one ONB of Hn, TA can be identified with a Hermitian ds x ds matrix. We fix

orthonormal bases {e^}^ of "Hat and introduce the random variables:

' " (

where * = {U N }, U N G U(d N ) = OBN N . We also define

1

1.21

(U N T£U N ef,ef)---TrT N
d N

(1.3)

Evidently,

T-itf (*) = t E ^-(*) = t- E ^(*) + °w ( L4 )

3=1 3=1

(where the o(l) term is independent of ^). Thus,

LEMMA 1.1. |Zll ISZ] The ergodic property of an ONB \I/ (SV) is equivalent to:

1 - 1

i im m E T Y »W = ' VA G *°( M ) • ( L5 )

jV->oc A/ ^— ' d n

n=l

As mentioned in the introduction, it follows by a standard diagonal argument that almost
all the individual elements (Aipsj^Nj) tend to oo{A) for all A. We do not discuss this step
since it is nothing new.

cS

STEVE ZELDITCH

1.3. Moment map interpretation. In the case where the components of A at are distinct,
the covex polytope V% is the permutahedron determined by A, that is, the simple convex

polytope defined as the convex hull of the points {a(Ajv)} where a £ SV runs over the
symmetric group on d^ letters (i.e. the Weyl group of U(cIn))- The center of mass is the
unique point X £ V? so that

Xa(X N ) = <=► X = — ^ V

where XY = X — Y is the vector from X to Y. The center of mass is evidently invariant
under Sd N , hence has the form (a, a, ... , a) for some a and clearly a = Ylj=i ^j-

In effect, we want to asymptotically calculate the moments of inertia of the sequence of
permutahedra associated to a Toeplitz operator.

Per mut ahedr on

1.4. Symmetric polynomials and Schur polynomials. The elementary symmetric poly-
nomials of N variables are defined by

e k (Xi, . . . ,X N ) = ^2 x h ---x iN .

h<i2<---<ik<N

If one replaces < by < one obtains the complete symmetric polynomials h^. The Schur
polynomials are symmetric polynomials defined by

S x = det [hxi+j-i) = det (e^ +j -i)

where [i is a dual partition to A.

1.5. Fourier transform of the orbit. We can compute the moments using the Fourier

transform (0.10) of the orbital measure on the orbit of -D(A).

An explicit formuae for fi\(X) is given in the first line of the proof of Theorem 5.1 of [OVJ:
Lemma 1.2. ForU(d),

Here, £(fi) is the number of rows of the partition fi. The degree of is the number of
boxes.

RANDOM ORTHONORMAL BASES OF SPACES OF HIGH DIMENSION 9

Since we would like to shift the center of mass of V r to the origin, we mainly consider
p>ji(X) the Fourier transform of the traceless orbit (see (0.9)).

2. Proof of Proposition [TJ Moment asymptotics
2.1. Second moment asymptotics. We now prove:

Lemma 3. [ZH [Z2l [SZ] Let A = (Ai, . . . , X dN ) G K djv , and Zei A)(AaO denote the trace zero
diagonal matrix with entries (0.9). Thus, Pi(A) = (0.9). Then

= / ^(^(WfcZC/ = , (2.1)

where as above, dll is the normalized Haar probability measure on U(aV).

This Lemma was proved in |Z11 IZ21 ISZj using the so-called Itzykson-Zuber-Harish-Chandra
formua for the Fourier transform of the orbit, and again using Gaussian integrals. The
proof we give here generalizes better to higher moments. We also sketch a proof using the
Weingarten formulae.

Proof. We use Lemma L2 to obtain

AS^S^iX)

E\\J dN (U*D(X)U)\\ 2 = (d N - 1)! • • • 0!

»M=2 M < dN 0* +dN ~ ^ + dN ~ 2 ) ! ' •

We sum over the Young diagrams with exactly two boxes and < d^ rows. There are just
two of them: one row of two boxes or two rows of one box each corresponding respectively
to the Schur functions 5(2,0) , £(1,1) ■ Note that S X k = is the kth elementary symmetric
function and S^) = hj. is the complete kth degree symmetric function.

We then translate A to A so that ^ ■ Aj = ei(A) = 0, i.e. we replace D(Ajv) by D (Xn)-
Since the degree \fj,\ = 2, then we can only use \i = (2), (11) and

0C i 0C j .

S(i, 1) = e 2 = ^2 XiX i> S (w) = e i ~ 62 = x ? + '

i<j j i<j

But

Ae 2 = 0, A{el-e 2 ) = 2\\Ve 1 \\ 2 = 2d N .

For each monomial X{Xj we have AXiXj = 25^. Thus, ASn t i\ = and AS (20) = 2dj^.
Since the Schur polynomials are homogeneous of degree 2, we can remove the % under the
Schur polynomials to get an overall factor of —1, which is cancelled by the — sign from A.
Thus,

E||J djv (^ (A)£/)|| 2 = (2^)(d^-l)!^§ = ^g^5 PW)) (<X)

= ra^o)(M = ^(e?-e 2 )(A).

Since ei(Ajv) = we find that

E||J djv (f/*A>(A)£/)|| 2 = -^re 2 (A) = ^p 2 (A

N)

10

STEVE ZELDITCH

Here we use that

e x = px, 2e 2 = e-iPx - p 2 .
The formula agrees with the one stated in the Lemma |3j

□

2.2. Weingarten formulae for the expectation. As a second proof, we use the Wein-
garten formula for integrals of polynomials over U(N) [W]. We denote the eigenvalues of
D (X) by A. Then,

||diag(£/*A)(A)f/)|| 2 = E hj2 ^AhJ:i\Uinm h \ 2 .

The Weingarten formulae for these special polynomials state that asymptotically \fd^\Uij\ 2
is a complex Gaussian random variable of mean zero and variance one. Thus, to leading
order,

Iu(d N ) \Uhoi \ 2 Whj2 \ 2 dU ~ d N (1 + 8j x

32)1

and

E^Ah^EiJuujUiuAm^dU ~ ^ 1 (2E i A? + E il ^A jl A i2 ). (2.2)

Since

we get

2.3. Proof of Proposition [TJ Variance and fourth moment asymptotics. We now

prove the 4th moment identity in Proposition [TJ which is the main new step in this article.

To calculate the variance of Y$ we use the expression in Lemma [l] in terms of jx^ and then
use the formula of Lemma 11.21

A Schur polynomial S nij _ }Jld (xi, . . . ,Xd) of degree n in d variables is parameterized by a
partition of of the degree n = n% + n 2 + • • • + rid into d parts. When n = 4 and d > 4 there
are 5 partitions:

^1,1,1,1 O^) ^4 ^^X<i<j<k<£ "^i^j^k^ti

S 2 ,i,i(xi, . . . ,x N ) = eie 3

5*2,2,0

= e\- eie 3

(2.3)

5*4,0,0

= e\ — 3efe 2 + 2eie 3 + e\

, 53,1,0

= e\e 2 - el - e x e z .

We note that Ae&(X) = for all k, so Ae^e™ = 2Ve/c ■ Ve n . Also, Vei is a constant vector.
So A 2 eie 3 = Vei • VAe 3 = and

A 2 ele 2 = 4A( ei Vei • Ve 2 ) = 8Ve! ■ V(Vei ■ Ve 2 ) = 8TrHesse 2 = 0.

RANDOM ORTHONORMAL BASES OF SPACES OF HIGH DIMENSION

11

Here, Hess denotes the Hessian. We also use that A(V/ • Vg) = 2Hess(/) • Hess(g) when
Af = Ag = 0. Also,

Vei • V(Vei • Ve 2 ) = (1, 1, . . . , 1) • ^

j.k

d 2 e 2 d
dxjdxk dxk

Tr Hess(e2) = 0.

Then,

A 2 e 2 2 = 2A(Ve 2 • Ve 2 ) = 4||Hess(e 2 )| | 2 = 4d N (d N - 1).

Further, Aef = 2Vei ■ Vei = 2d,N, so that

A 2 e\ = A{2(Ae 2 1 )e 2 1 + 2Ve 2 ■ Ve 2 ) = A{U N e\ + 2e\d N ) = 12d%.
We recall Newton's identities,

' ei =

= Pi

2e 2

= eiPi

-P2

3e 3

= e 2 pi

- eiP2 + P3

k 4e 4

= em

- e 2 p 2 + eip 3 - Pa

We note that Ae^ = for all k, so at X = 0,

= L

AS 2j1j1 =

A 2

< A 2 S 2:2 fl =

-- A

AS^Qfi =

A 2

in

k A5 3 ,i, = 8V ei • V(V ei • Ve 2 ) - ||Hess(e 2 )|| 2 = -4||Hess(e 2 )|| 2 = -4d N (d N - 1).

(2.4)

By routine calculations and Lemma 1.2 we have

A 2 Aa(()) :

A 2 5 M (0)5 M (»A)

- (d N l)!-"0! ^1^=4 ( Atl+djv _i)!( A12+ d JV -2)!... Al£ijv !

= (d N -l)l(d N - 2)1^0^0^1

I (J 1M A 2 54.o,o(0)54,o,o(^A)

+ l d iV-lJ! ( djv+3 )!

_i_ (A \\\( A qm A 2 53,i,o(0)53,i, (iA)

+ [d N - i)l{d N - zy. (djv+2)!(djv _ 1)!

A 2 52,2,0(0)S2,2,0(»A)

(djv + l)^(<ijv-l)

A 2 5 4 ,o,o(0)54,o,o(»A)
(d JV +3)(d J v+2)(d J v+l)d JV ' (d N +2)(d N +l)d N (d N -2)-

(2.5)

+

+

A 2 S 3 ,i,o(0)g3.i,o(iA)

12

STEVE ZELDITCH

By (2.4), we then have

A 2 Aa(0)

4d N (d N -l)S2,2,o(iA-)
(d N +l)d%(d N -l)

Recalling (2.3) and that ei(A)

(12d 2 N +2d N (d N -l))Si, , (iA)
+ {d N +3)(d N +2)(d N +l)d N +

= 0, we get

-4d N (d N -l)S 3 ,i,o(iA)
(d N +2)(d N +l)d N (d N -2)-

(2.6)

A 2 Aa(0)

(d N +l)d 2 N (d N -l)

Further recalling that 2e2 = eipi

Ad N (d N -l)e 2 (iA) (l2d 2 N +id N (d N -l))el{itL)

+

A 2 ^(0)

d N {d N -l)pl(i\)

(d N +3)(d N +2){d N +l)d N

- P2 we finally get

(3d 2 N +d N (d N -l))p 2 (iA)

+

—4d N (d N — l)e^(iK)
(d N +2)(d N + l)d N (d N -2)-

+

+

-d N (d N -l)p 2 (iA)

(2.7)

{21

(d N +l)d^(d N -l) 1 (d N +3)(d N +2)(d N +l)d N 1 (d JV +2)(<i J v+l) ( i i v(d J v-2)

Since the polynomials are homogeneous of degree 4, the factor of i inside the polynomials
may be removed, and we get

(3d 2 N +d N (d N -l))p 2 (A)

+

-d N (d N -l)p%(A)

FN T (TJ*D CAVnil 4 = d N (d N -l)p 2 (A)

JE 'IKdjvV L/ ^OW^II - (diV+ i )d 2 r((ijv _ 1) -r (diV+3)(djv+2)(diV+1)diV i (^+2)^+1)^(^-2) ■
As A" — > oo the leading asymptotics of the outer terms cancel and the middle term is

We note that P2 f, Ajv - > is bounded. If the the empirical measure of

asymptotic to 4rp\{k

N

eigenvalues tends to a limit measure, then

d N
P2(Ajy)

tends to its second moment.

Together with Lemma 2.1, this completes the proof of Proposition [TJ Corollary [T] follows
by subtracting the square of the expectation.

3. Completion of proof of Theorem \T\

By the assumption of Definition (0.7)

u(A)

d

N

-Tr 7^ + 0(1)

(3.1)

By Lemma TlCorollary

N

arc

the variances of the independent random variables j^Y 1
bounded. Hence, as explained in the introduction (see also \Z1\ ISZj ). (1.5) follows from

Lemma [T]-Corollary [T] and the Kolmogorov strong law of large numbers, which gives

lim ,

1 N

-Y

n=l

d n

almost surely .

(3-2)

By (1.4)

sup \X

ONB N

d

-Y,

N

N

0(1)

Hence also

lim ,

N^oo N

1 N

v

n=l

1 X A

almost surely

(3.3)

If the dimensions d^ grow fast enough so that ^— is summable, then we obtain a stronger
form from the fact that X^i^^r^n 4 i s finite hence the general term must tend to zero
almost everywhere. It follows again that E^-A"^ 1 — > almost everywhere.

RANDOM ORTHONORMAL BASES OF SPACES OF HIGH DIMENSION

13

4. Applications

4.1. Fat tori. Theorem [l] applies to eigenspaces of the Laplacian on the flat torus ]R d /Z d (or
other rational lattices) of dimension > 5 and for many eigenspaces in dimensions d = 2,3, 4.

Proposition 4.1. Random orthornomal bases of A- eigenspaces of the flat torus M, d /7j d are
quantum ergodic for d > 5. Also for d = 2, 3, 4 for special eigenspaces (specified below).

The only condition on the eigenspaces for Theorem [T] is that (0.7) holds, and we now recall

the known results on this problem. Given A e we denote the eigenspaces on a flat torus,
enumerated in order of the eigenvalue by Hn and by II jv the orthogonal projection to 1-Lm q

Lemma 4.2. The condition (0.7) is valid in dimensions > 5 on M~/'L d . That is,

1 f

-j—Tr AUn ~ / a(x, u)dx A du.

djy J s*T m

It follows that j-Y£ —> almost surely.

In dimensions 2, resp. 3, resp. 4 there are restrictions on the sequence of eigenvalues

n/ Z n

—TrTl N A= / aA(x,k)dx.

given in |EH] . resp. [DSP], resp. [P]. For eigenvalues in the allowed sequences, (0.7) is
valid.

Proof. We use the basis = e^ k,x ' with \k\ = /ijv- Then

( Ae k> e k) = [ a A (x,k)dx.

Hence

d*N , ,T~~ JR"/Z n

k:\k\=n N 1

In dimensions n > 5, djv ~ I^n~ 2 - It is proved that lattice points of fixed norm on a sphere
of radius y/n become uniformly distributed as n — > oo [P] . It follows that

— / o~a{x, k)dx — ¥ I a(x, u)dx A dco.

N .iii JR"/Z n Js*T m
k:\k\=ti N I

As in the last step of the proof of Theorem jlj Kj^X^ is summable when n > 5.

The Liouville limit formula is true in dimension 4 when the number of lattice points grows
linearly in n. The condition on n is given in [Pj. In dimension 3, the equidistribution result
is proved in |DSPj with similar conditions on the sequence of integers n.

Dimension 2 is more complicated. In dimension 2, the eigenvalues of integers n for which
there exist lattice points (a, b) on the circle a 2 + b 2 = n. It is necessary that all prime factors
of n are congruent to 1 modulo 4. In [EH] it is shown that for almost all such n, the lattice
points on the circle become uniformly distributed as n — ¥ oo. □

Remark: In the case of a generic lattice L C M. d , the multiplicity of eigenvalues of A on IR d / L
is two. The analogue of the eigenspaces above are spectral subspaces for a/A of shriking width
w. Thus, one considers the exponentials e 1 ^'^ for £ G L with \t\ e [A — Cw, X + w]. It follows
from the lattice point results of [G] that in dimensions d > 5, the number of eigenvalues of

thanks to Z. Rudnick for explanations and references

14

STEVE ZELDITCH

an irrational flat torus in [A, A + 0(X x ] is of order X d 2 . The question whether the trace

asymptotics (0.7) hold for the span of the corresponding eigenf unctions does not appear to

have been studied.

4.2. Quasi-modes. Theorem [T] is not restricted to eigenspaces of the Laplacian and is
equally valid for spaces of quasi-modes. We refer to |CVt IPo] for background on quasi-
modes. Following |Poj . we define a C°° quasimode of infinite order for h 2 A with index set
A4h to be a family

Q = {(lPm(-,fr),Vrn{fr)) ■ m G M h }

of approximate eigenfunctions satisfying

{i)\\{h 2 A- ^ m {n))^ m {-M\Hs = M (h M ), (vmgZ+),

(4.1)

m^ m ^ n ) ~ 6 mn \ = M (h M ), (VM G Z+).

It follows by the spectral theorem that for any M G Z + , there exists at least one eigenvalue
of h 2 A in the interval

i r U = W{h) - h M ^ m {h) + h M l

and

\\E IhA ^ k -^ k \\ Hs =0 M (h M ). (4.2)

Here, Ej denotes the spectral projection for h 2 A corresponding to the interval /. We denote
the quasi-classical eigenvalue spectrum of hyAbj

QSph = {^ m {h) : m G Mn}-

Since quasi-eigenvalues /j, m (h) are only defined up to errors of order h°°, there is a notion
of 'multiple quasi-eigenvalue' defined as follows: we say fi m (h) ~ Unifi) if A*m — A*n — 0(h°°)
and define the multiplicity of ix m (h) by

mult{ji m {h)) = #{n : fx m (h) ~ /i n (^)} = dim Span{^ n (-, ti) : (^ 2 A - fi m (h))^ n = 0(h°°)}.
We then introduce slightly larger intervals I m ,n (if need be) so that

QSp(h) C [J X mjfi , Z m>h nl n>h = {m^n).

m£M'

Here, .A/f consists of equivalence classes of indices (corresponding to equivalence classes of
quasimodes). We denote by the span of the quasimodes {ip m (-,h) : fJ, m (h) G 2^}. Then

||^-^|| = 0(0, if vent-

Theorem [l] applies to quasi-mode spaces "H^ as long as their dimensions tend to infinity and
as long as there exists a unique limit state for dir ^ nh TrA\ n r^. One might expect true modes

(eigenfunctions) with eigenvalues in the intervals to be close to linear combinations of
the quasi-modes with quasi-eigenvalues in that interval. The question raised by Theorem [T]
is whether they behave like random linear combinations or not. If they do, Theorem [T] gives
their quantum limits.

In particular, this bears on the question whether A-eigenf unctions of compact Riemannian
manifolds (M, g) with KAM geodesic flow might be quantum ergodic. It seems unlikely that
they are, but we are not aware of a proof that they are not. For such KAM (M,g), a

RANDOM ORTHONORMAL BASES OF SPACES OF HIGH DIMENSION

15

large family of quasi-modes is constructed in \CV\ IPoj which localize on the invariant tori
of the KAM Cantor set of tori. Without reviewing the results in detail, the 'large' family
has positive spectral density i.e. the number of quasi-eigenvalues < /i grows like a positive
constant times /i n where n = dim M.

To our knowledge, the multiplicities and trace asymptotics for KAM quasi-modes have
not been studied at this time. As in the discussion of flat tori, one would need to determine
the equidistribution law of the tori in the invariant Cantor set corresponding to eigenvalues
(or pseudo-eigenvalues) of \J~K in very short intervals I\ = [A — w, A + wj. The orthonormal
basis of eigenfunctions is not simple to relate to the near orthonormal basis of quasi-modes in
this case, but we might expect that a positive density of the eigenfunctions are mainly given
as linear combinations of KAM quasi-modes with quasi-eigenvalues very close to the true
eigenvalues. Whether or not they are quantum ergodic would reflect the extent to which they
are sufficiently random combinations of quasi-modes and the extent to which the collection
of quasi-modes in 1^ is Liouville distributed.

REFERENCES

[BL] N. Burq and G. Lebeau, Injections de Sobolev probabilistes et applications, arXiv:1111.7310|

[CV] Y.Colin de Verdiere, Quasi-modes sur les varietes Riemanniennes compactes, Invent. Math. 43
(1977), 15-52.

[DSP] W. Duke and R. Schulze-Pillot, Representation of integers by positive ternary quadratic forms and

equidistribution of lattice points on ellipsoids. Invent. Math. 99 (1990), no. 1, 49-57.
[EH] P. Erdos and R. R. P. Hall, On the angular distribution of Gaussian integers with fixed norm. Paul

Erdos memorial collection. Discrete Math. 200 (1999), no. 1-3, 87-94.
[OV] G. Olshanski and A. Vershik, Ergodic unitarily invariant measures on the space of infinite Hermitian

matrices. Contemporary mathematical physics, 137 - 175, Amer. Math. Soc. Transl. Ser. 2, 175,

Amer. Math. Soc, Providence, RI, 1996.
[G] F. Gotze, Lattice point problems and values of quadratic forms. Invent. Math. 157 (2004), no. 1,

195-226.

[P] C. Pommerenke, Uber die Gleichverteilung von Gitterpunkten auf m-dimcnsionalen Ellipsoiden.

Acta Arith. 5 1959 227-257.
[Po] G. Popov, Invariant tori, effective stability and quasimodes with exponentially small errors I-

Birkhoff normal forms, Ann. Henri Poincare 1 (2000), 223-248.
[S] S. Samuel, U(N) integrals, 1/N, and the De Wit-'t Hooft anomalies. J. Math. Phys. 21 (1980), no.

12, 2695- 2703.

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

positive line bundles. Comm. Math. Phys. 200 (1999), no. 3, 661-683.
[W] D. Weingarten, Asymptotic behavior of group integrals in the limit of infinite rank. J. Math. Phys.

19(5), 999-1001 (1978).

[Zl] S. Zelditch, Quantum ergodicity on the sphere. Comm. Math. Phys. 146 (1992), no. 1, 61-71.
[Z2] S. Zelditch, A random matrix model for quantum mixing. Internat. Math. Res. Notices 1996, no.
3, 115-137.

[Z3] S. Zelditch, Large N limits of coadjoint orbits (in preparation).

Department of Mathematics, Northwestern University, Evanston, IL 60208-2370, USA
E-mail address: zelditch@math.northwestern.edu