Essentially normal operators

Kenneth R. Davidson

Abstract. This is a survey of essentially normal operators and related de-

velopments. There is an overview of Weyl–von Neumann theorems about
expressing normal operators as diagonal plus compact operators. Then we
consider the Brown–Douglas–Fillmore theorem classifying essentially normal
operators. Finally we discuss almost commuting matrices, and how they were
used to obtain two other proofs of the BDF theorem.
Mathematics Subject Classification (2000). 47-02,47B15,46L80.
Keywords. essentially normal operator, compact perturbation, normal, diag-
onal, almost commuting matrices, extensions of C*-algebras.

1. Introduction
Problem 4 of Halmos’s Ten problems in Hilbert space [26] asked whether every
normal operator is the sum of a diagonal operator and a compact operator. I believe
that this was the first of the ten problems to be solved. Indeed two solutions were
independently produced by Berg [6] and Sikonia [39] almost immediately after
dissemination of the question. But that is only the beginning of the story, as, like
many of Paul’s problems, the answer to the question is just a small step in the
bigger picture.
The subsequent discussion in Halmos’s article goes in several directions. In
particular, he discusses operators of the form normal plus compact. It is not at all
clear, a priori, if this set is even norm closed. It turned out that the characterization
of this class by other invariants is very interesting.
It is immediately clear that if T is normal plus compact, then T ∗ T − T T ∗ is
compact. Operators with this latter property are called essentially normal. Not all
essentially normal operators are normal plus compact. For example, the unilateral
shift S acts on a basis {en : n ≥ 0} by Sen = en+1 . It evidently satisfies
S ∗ S − SS ∗ = I − SS ∗ = e0 e∗0 ,

Partially supported by an NSERC grant.

2 K.R.Davidson

which is the rank one projection onto Ce0 . Thus the unilateral shift is essentially
normal. However it has non-zero Fredholm index:

ind S = dim ker S − dim ker S ∗ = −1.

For a normal operator N , one has ker N = ker N ∗ . Thus if N is Fredholm, then
ind N = 0. Index is invariant under compact perturbations, so the same persists
for normal plus compact operators.
Brown, Douglas and Fillmore [9] raised the question of classifying essentially
normal operators. The answer took them from a naive question in operator theory
to the employment of new techniques from algebraic topology in the study of
C*-algebras. They provided a striking answer to the question of which essentially
normal operators are normal plus compact. They are precisely those essentially
normal operators with the property that ind (T − λI) = 0 for every λ 6∈ σe (T ),
namely the index is zero whenever it is defined. This implies, in particular, that
the set of normal plus compact operators is norm closed.
Had they stopped there, BDF might have remained ‘just’ a tour de force that
solved an interesting question in operator theory. However they recognized that
their methods had deeper implications about the connection between topology
and operator algebras. They defined an invariant Ext(A) for any C*-algebra A,
and determined nice functorial properties of this object in the case of separable,
commutative C*-algebras. They showed that Ext has a natural pairing with the
topological K-theory of Atiyah [4] which makes Ext a K-homology theory. This
opened up a whole new world for C*-algebraists, and a new game was afoot.
At roughly the same time, George Elliott [18] introduced a complete algebraic
invariant for AF C*-algebras. These C*-algebras, introduced by Bratteli [8], are
defined by the property that they are the closure of an (increasing) union of finite
dimensional sub-algebras. It was soon recognized [19] that this new invariant is
the K0 functor from ring theory. A very striking converse to Elliott’s theorem
was found by Effros, Handelmann and Shen [17] which characterized those groups
which arise as the K0 group of an AF algebra.
The upshot was that two very different results almost simultaneously seeded
the subject of C*-algebras with two new topological tools that provide interesting
new invariants, namely Ext and K0 . These results created tremendous excitement,
and launched a program which continues to this day to classify amenable C*-
algebras. It revitalized the subject, and has led to a sophisticated set of tools which
describe and distinguish many new algebras. It is fair to say that the renaissance of
C*-algebras was due to these two developments. Indeed, not only are they related
by the spirit of K-theory, they are in fact two sides of the same coin. Kasparov
[30] introduced his bivariant KK-theory shortly afterwards which incorporates the
two theories into one.
It is not my intention to survey the vast literature in C*-algebras which
has developed as a consequence of the introduction of K-theory. I mention it to
highlight the fallout of the pursuit of a natural problem in operator theory by three
 Essentially normal operators 3

very insightful investigators. I will limit the balance of this article to the original
operator theory questions, which have a lot of interest in their own right.

2. Weyl-von Neumann Theorems

In 1909, Hermann Weyl [46] proved that every self-adjoint bounded operator A
on a separable Hilbert space can be written as A = D + K where D is a diago-
nal operator with respect to some orthonormal basis and K is compact. Hilbert’s
spectral theorem for Hermitian operators says, as formulated by Halmos [27], that
every Hermitian operator on Hilbert space is unitarily equivalent to a multiplica-
tion operator Mϕ on L2 (µ), where ϕ is a bounded real-valued measurable function.
The starting point of a proof of Weyl’s theorem
is the observation that if f is any
function in L2 (µ) supported on ϕ−1 [t, t + ε) , then f is almost an eigenvector
in the sense that kMϕ f − tf k < εkf k. One can carefully extract an orthonormal
basis for L2 (µ) consisting of functions with increasingly narrow support.
To make this more precise, suppose for convenience that A has a cyclic vector.
Then the spectral theorem produces a measure µ on the spectrum σ(A) ⊂ R so
that A is identified with Mx on L2 (µ). Let Pn be the span of the characteristic
functions of diadic intervals of length 2−n . Then the previous observation can be
used to show that kPn A − APn k < 2−n . So a routine calculation shows that A is
approximated within 21−n by the operator
X
Dn = Pn APn + (Pk+1 − Pk )A(Pk+1 − Pk ),
k≥n

and A − Dn is compact. The operator Dn is a direct sum of finite rank self-adjoint

operators, and so is diagonalizable—providing the desired approximant.
Weyl observed that if A and B are two Hermitian operators such that A − B
is compact, then the limit points of the spectrum of A and B must be the same. We
now interpret this by saying that the essential spectra are equal, σe (A) = σe (B),
where σe (·) denotes the spectrum of the image in the Calkin algebra B(H)/K.
Von Neumann [45] established the converse: if two Hermitian operators have the
same essential spectrum, then they are unitarily equivalent modulo a compact
perturbation.
Halmos’s questions asks for an extension to normal operators. It seems to
require a new approach, because the trick of compressing a Hermitian operator to
the range of an almost commuting projection Pk+1 −Pk yields a Hermitian matrix,
but the same argument fails for normal operators. Also the spectrum is now a
subset of the plane. David Berg [6] nevertheless answered Halmos’s question by
adapting this method. Sikonia gave a similar proof at the same time in his doctoral
thesis (see [39]). Other proofs came quickly afterwards (for example Halmos [27]).
A proof that does the job simultaneously for a countable family of commuting
Hermitian operators {Ai } works by building a single Hermitian operator A so
that C∗ (A) contains every Ai . To accomplish this, first observe that the spectral
theorem shows that C∗ ({Ai }) is contained in a commutative C*-algebra spanned
4 K.R.Davidson

Pfamily {Ej } of commuting projections. Consider the Hermitian

by a countable
operator A = j≥1 3−j Ej . It is easy to see that
1
3 E1 ≤ E1 A ≤ 12 E1 and 0 ≤ E1⊥ A ≤ 16 E1⊥ .
Thus it follows that the spectrum of A is contained in [0, 1/6] ∪ [1/3, 1/2], and that
the spectral projection for [1/3, 1/2] is E1 . Similarly, each projection Ej belongs
to C∗ (A). It follows that each Ai is a continuous function of A. Diagonalizing A
modulo compacts then does the same for each Ai , although one cannot control the
norm of the compact perturbations for more than a finite number at a time.
At this point, we introduce a few definitions to aid in the discussion.
Definition 2.1. Let J be a normed ideal of B(H). Two operators A and B are
unitarily equivalent modulo J if there is a unitary U so that A − U BU ∗ ∈ J. We
say that A and B are approximately unitarily equivalent modulo J if there is a
sequence of unitaries Uk so that A − Uk BUk∗ ∈ J and
lim kA − Uk BUk∗ kJ = 0.
k→∞
We write A ∼J B. When J = B(H), we simply say that A and B are approximately
unitarily equivalent and write A ∼a B.
A careful look at Weyl’s proof shows that the perturbation will lie in certain
smaller ideals, the Schatten classes Sp with norm kKkp = Tr(|K|p )1/p , provided
that p > 1; and this norm can also be made arbitrarily small. Kuroda [32] improved
on this to show that one can obtain a small perturbation in any unitarily invariant
ideal J that strictly contains the trace class operators. So any Hermitian operator
is approximately unitarily equivalent modulo J to a diagonalizable operator. Berg’s
proof works in a similar way via a process of dividing up the plane, and actually
yields small perturbations in Sp for p > 2. Likewise, for a commuting n-tuple, one
can obtain small perturbations in Sp for p > n. It is natural to ask whether this
is sharp. Elementary examples show that one cannot obtain perturbations in Sp
when p < n.
In the case of n = 1, there is an obstruction found by Kato [31] and Rosen-
blum [37]. If the spectral measure of A is not singular with respect to Lebesgue
measure, then there is no trace class perturbation which is diagonal. In particular,
Mx on L2 (0, 1) is such an operator. For n ≥ 2, Voiculescu [41, 42] showed that
every commuting n-tuple of Hermitian operators is approximately unitarily equiv-
alent to a diagonalizable n-tuple modulo the Schatten class Sn . (See [14] for an
elementary argument.) Moreover, Voiculescu identified a somewhat smaller ideal
S− n which provides an obstruction when the n-tuple has a spectral measure that is
not singular with respect to Lebesgue measure on Rn . Bercovici and Voiculescu [5]
strengthened this to the analogue of Kuroda’s theorem, showing that if a unitar-
ily invariant ideal is not included in S−n , then a small perturbation to a diagonal
operator is possible.
The ideas involved in Voiculescu’s work mentioned above build on a very
important theorem of his that preceded these results, and had a direct bearing on
 Essentially normal operators 5

the work of Brown, Douglas and Fillmore and on subsequent developments in C*-
algebras. This is known as Voiculescu’s Weyl-von Neumann Theorem [40]. Rather
than state it in full generality, we concentrate on some of its major corollaries.
The definition of approximate unitary equivalence can readily be extended to two
maps from a C*-algebra A into B(H): say ρ ∼a σ for two maps ρ and σ if there is
a sequence of unitary operators Un such that
lim kρ(a)Un − Un σ(a)k = 0 for all a ∈ A.
n→∞
One similarly defines approximate unitary equivalence relative to an ideal.
Voiculescu showed that if A is a separable C*-algebra acting on H and ρ is a
∗-representation which annihilates A∩K, then id ∼K id⊕ρ, where id is the identity
representation of A. Let π denote the quotient map of B(H) onto the Calkin algebra
B(H)/K. Voiculescu’s result extends to show for two representations ρ1 and ρ2 , one
has ρ1 ∼a ρ2 if and only if ρ1 ∼K ρ2 . In particular, this holds provided that
ker ρ1 = ker πρ1 = ker ρ2 = ker πρ2 .
If N is normal, then it may have countably many eigenvalues of finite multi-
plicity which do not lie in the essential spectrum. However they must be asymptot-
ically close to the essential spectrum. One can peel off a finite diagonalizable sum-
mand, and make a small compact perturbation on the remainder to move the other
eigenvalues into σe (N ). One now has a normal operator N 0 with σ(N 0 ) = σe (N 0 ). If
one applies Voiculescu’s Theorem to C∗ (N 0 ), one recovers the Weyl-von Neumann-
Berg Theorem.
Another consequence of this theorem was a solution to Problem 8 of Halmos’s
ten problems, which asked whether the reducible operators are dense in B(H). One
merely takes any representation ρ of C∗ (T ) which factors through C∗ (T ) + K/K
to obtain T ∼a T ⊕ ρ(T ). Moreover, one can show that id ∼K σ where σ is a
countable direct sum of irreducible representations.
The implications of Voiculescu’s theorem for essentially normal operators will
be considered in the next section. We mention here that a very insightful treatment
of Voiculescu’s theorem is contained in Arveson’s paper [3]. In particular, it pro-
vides a strengthening of the results for normal operators. Hadwin [25] contains a
further refinement which shows that ρ1 ∼K ρ2 if and only if rank ρ1 (a) = rank ρ2 (a)
for all a ∈ A. All of these ideas are treated in Chapter 2 of [15].

3. Essentially normal operators

We return to the problem of classifying essentially normal operators. Let T be
essentially normal. Then t = π(T ) is a normal element of the Calkin algebra. So
C∗ (t) ' C(X) where X = σ(t) = σe (T ). This determines a ∗-monomorphism τ
of C(X) into B(H)/K determined by τ (z) = t. Evidently τ determines T up to a
compact perturbation. Two essentially normal operators T1 and T2 are unitarily
equivalent modulo K if and only if σe (T1 ) = σe (T2 ) =: X and the associated
monomorphisms τ1 and τ2 of C(X) are strongly unitarily equivalent, meaning
6 K.R.Davidson

that there is a unitary U so that ad π(U )τ1 = τ2 . (The weak version would allow
equivalence by a unitary in the Calkin algebra. This turns out to be equivalent
for commutative C*-algebras.) A monomorphism τ is associated to an extension
of the compact operators. Let E = π −1 (τ (C(X)) = C∗ (T ) + K. Then τ −1 π is a
∗-homomorphism of the C*-algebra E onto C(X) with kernel K. We obtain the
short exact sequence
0 → K → E → C(X) → 0.

For example, the unilateral shift S is unitarily equivalent to the Toeplitz

operator of multiplication by z on H 2 . So one readily sees that C∗ (S) is unitarily
equivalent to the Toeplitz C*-algebra

T(T) = {Tf + K : f ∈ C(T) and K ∈ K}.

The map τ1 (f ) = π(Tf ) is a monomorphism of C(T) into the Calkin algebra. It

is not hard to use the Fredholm index argument to show that this extension does
not split; i.e. there is no ∗-monomorphism of C(T) into T taking z to a compact
perturbation of Tz . Therefore this extension is not equivalent to the representation
of C(T) on L2 (T) given by τ2 (f ) = π(Mf ) where Mf is the multiplication operator.
Turning the problem around, Brown, Douglas and Fillmore consider the class
of all ∗-monomorphisms of C(X) into B(H)/K for any compact metric space
X; or equivalently they consider all extensions of K by C(X). Two extensions
are called equivalent if the corresponding monomorphisms are strongly unitarily
equivalent. The collection of all equivalence classes of extensions of C(X) is de-
noted by Ext(X). One can turn this into a commutative semigroup by defining
[τ1 ] + [τ2 ] = [τ1 ⊕ τ2 ], which uses the fact that we can identify the direct sum H ⊕ H
of two separable Hilbert spaces with the original space H.
More generally, one can define Ext(A) for any C*-algebra. The theory works
best if one sticks to separable C*-algebras. Among these, things work out partic-
ularly well when A is nuclear.
The Weyl–von Neumann–Berg Theorem is exactly what is needed to show
that Ext(X) has a zero element. An extension is trivial if it splits, i.e. there is
a ∗-monomorphism σ of C(X) into B(H) so that τ = πσ. The generators of
σ(C(X)) can be perturbed by compact operators to commuting diagonal operators.
The converse of von Neumann is adapted to show that any two trivial extensions
are equivalent. An elementary argument can be used to construct approximate
eigenvectors. Repeated application yields τ ∼K τ ⊕σ, where σ is a trivial extension.
So the equivalence class of all trivial extensions forms a zero element for Ext(X).
In fact, Ext(X) is a group. Brown, Douglas and Fillmore gave a complicated
proof, which required a number of topological lemmas. The proof was significantly
simplified by Arveson [2] by pointing out the crucial role of completely positive
maps. The map τ may be lifted to a completely positive unital map σ into B(H),
meaning that τ = πσ. Then the Naimark dilation theorem dilates this map to a
 Essentially normal operators 7

∗-representation ρ of C(X) on a larger space K ⊃ H; say

 
τ (f ) ρ12 (f )
ρ(f ) = .
ρ21 (f ) ρ22 (f )
Since the range of τ commutes modulo compacts, it is not hard to see that the
ranges of ρ12 and ρ21 consist of compact operators. It follows that πρ22 is a ∗-
homomorphism of C(X). The map πρ22 ⊕ τ0 , where τ0 is any trivial extension,
yields an inverse.
More generally, Choi and Effros [12] showed that Ext(A) is a group whenever
A is a separable nuclear C*-algebra. The argument uses nuclearity and the struc-
ture of completely positive maps to accomplish the lifting. The dilation follows
from Stinespring’s theorem for completely positive maps. Voiculescu’s Theorem
provides the zero element consisting of the class of trivial extensions. (See Arveson
[3] where all of this is put together nicely.) When A is not nuclear, Anderson [1]
showed that Ext(A) is generally not a group. Recently Haagerup and Thorbjornsen
[24] have shown that Ext of the reduced C*-algebra of the free group F2 is not a
group.
Next we observe that Ext is a covariant functor from the category of compact
metric spaces with continuous maps into the category of abelian groups. Suppose
that p : X → Y is a continuous map between compact metric spaces, and τ is an
extension of C(X). Build an extension of C(Y ) by fixing a trivial extension σ0 in
Ext(Y ) and defining
σ(f ) = τ (f ◦ p) ⊕ σ0 (f ) for all f ∈ C(Y ).
So far, we have seen little topology, although the original BDF proof used
more topological methods to establish these facts. Now we discuss some of those
aspects which are important for developing Ext as a homology theory. If A is a
closed subset of X, j is the inclusion map and p is the quotient map of X onto
X/A, then
j∗ p∗
Ext(A) −→ Ext(X) −→ Ext(X/A)
is exact. Ext also behaves well with respect to projective limits of spaces. If Xn
are compact metricQ spaces and pn : Xn+1 → Xn for n ≥ 1, define X = proj lim Xn
to be the subset of n≥1 Xn consisting of sequences (xn ) such that pn (xn+1 ) = xn
for all n ≥ 1. There are canonical maps qn : X → Xn so that qn = pn qn+1 . One
can likewise define proj lim Ext(Xn ). Since qn∗ defines a compatible sequence of
homomorphisms of Ext(X) into Ext(Xn ), one obtains a natural map
κ : Ext(proj lim Xn ) −→ proj lim Ext(Xn ).
The key fact is that this map is always surjective. Moreover, it is an isomorphism
when each Xn is a finite set. This latter fact follows by noting that when each Xn
is finite, X is totally disconnected. From our discussion of the Weyl–von Neumann
Theorem, C(X) is generated by a single self-adjoint element, and every extension
is diagonalizable and hence trivial. So Ext(X) = {0}.
8 K.R.Davidson

In [10], it is shown that any covariant functor from compact metric spaces
into abelian groups satisfying the properties established in the previous paragraph
is a homotopy invariant. That is, if f and g are homotopic maps from X to Y ,
then f∗ = g∗ . In particular, if X is contractible, then Ext(X) = {0}.
There is a pairing between Ext(X) and K 1 (X) which yields a map into the in-
tegers based on Fredholm index. First consider the group π 1 (X) = C(X)−1 /C(X)−1 0 .
An element [f ] of π 1 (X) is a homotopy class of invertible functions on X. Thus if
[τ ] ∈ Ext(X), one can compute ind τ (f ) independent of the choice of representa-
tives. Moreover, this determines a homomorphism γ[τ ] of π 1 (X) into Z. Hence we
have defined a homomorphism

γ : Ext(X) → Hom(π 1 (X), Z).

Similarly, by considering the induced monomorphisms of Mn (C(X)) into

B(H(n) )/K, we can define an analogous map

γ n : Ext(X) → Hom(GLn (X)/ GLn (X)0 , Z).

The direct limit of the sequence of groups GLn (X)/ GLn (X)0 is called K 1 (X).
The inductive limit of γ n is a homomorphism γ ∞ : Ext(X) → Hom(K 1 (X), Z).
This is the index map, and is the first pairing of Ext with K-theory.
When X is a planar set, an elementary argument shows that π 1 (X) is a
free abelian group with generators [z − λi ] where one chooses a point λi in each
bounded component of C \ X. An extension τ is determined by the essentially
normal element t = τ (z) ∈ B(H)/K. The index map is given by

γ([τ ])([z − λi ]) = ind (t − λi ).

The Brown–Douglas–Fillmore Theorem classifies essentially normal operators by

showing that when X is planar, the map γ is an isomorphism. The hard part of
the proof is a lemma that shows that when γ[τ ] = 0, one can cut the spectrum in
half by a straight line, and split τ as the sum of two elements coming from Ext of
the two halves. Repeated bisection eventually eliminates all of the holes.
An immediate consequence is the fact that an essentially normal operator
T is normal plus compact if and only if ind (T − λI) = 0 whenever λ 6∈ σe (T ).
More generally, two essentially normal operators T1 and T2 are unitarily equivalent
modulo K if and only if σe (T1 ) = σe (T2 ) =: X and

ind (T1 − λI) = ind (T2 − λI) for all λ ∈ C \ X.

Another immediate corollary is that the set of normal plus compact operators is
norm closed, since the essentially normal operators are closed, the set of Fredholm
operators is open, and index is continuous.
More information on the BDF theory with an emphasis on the K-theoretical
aspects is contained in the monograph [16] by Douglas. Most of the results men-
tioned here are treated in Chapter 9 of [15].
 Essentially normal operators 9

4. Almost commuting matrices

Peter Rosenthal [38] asked whether nearly commuting matrices are close to com-
muting. To make sense of this question, one says that A and B nearly commute if
kAB − BAk is small, while close to commuting means that there are commuting
matrices A0 and B 0 with kA − A0 k and kB − B 0 k both small. He makes it clear
that to be an interesting problem, one must obtain estimates independent of the
dimension of the space on which the matrices act. We may also limit A and B to
have norm at most one. Peter recalled, in a private communication, that he dis-
cussed this problem with Paul Halmos when he was his student at the University
of Michigan. The most interesting case, and the hardest, occurs when the matrices
are all required to be Hermitian. Halmos mentions the problem specifically in this
form in [28].
This ‘finite dimensional’ problem is closely linked to the Brown–Douglas–
Fillmore Theorem, as we shall see. I put finite dimensional in quotes because
problems about matrices which ask for quantitative answers independent of di-
mension are really infinite dimensional problems, and can generally be stated in
terms of the compact operators rather than matrices of arbitrary size.
If A and B are Hermitian matrices in Mn , then T = A + iB is a matrix
satisfying
[T ∗ , T ] = T ∗ T − T T ∗ = 2i(AB − BA).
So if A and B almost commute, then T is almost normal; and they are close to
commuting if and only if T is close to a normal matrix.
Halmos [26] defines an operator T to be quasidiagonal if there is a sequence
Pn of finite rank projections increasing to the identity so that kPn T −T Pn k goes to
0. The quasidiagonal operators form a closed set which is also closed under compact
perturbations. Normal operators are quasidiagonal, and thus so are normal plus
compact operators. Fredholm quasidiagonal operators have index 0.
Now suppose that T is an essentially normal operator which is quasidiagonal.
ThenP one can construct a sequence of projections Pn increasing to the identity so
that n≥1 kPn T − T Pn k is small. A small compact perturbation of T yields the
P
operator n≥1 ⊕Tn where
Tn = (Pn − Pn−1 )T (Pn − Pn−1 )|(Pn −Pn−1 )H .
The essentially normal property means that limn k[Tn∗ , Tn ]k = 0. So a positive
solution to the nearly commuting problem would show that T can be perturbed
by a block diagonal compact operator to a direct sum of normal operators. This
provides a direct link to the BDF theorem.
If one fixes the dimension n and limits A and B to the (compact) unit ball,
then a compactness argument establishes the existence of a function δ(ε, n) such
that if A and B are in the unit ball of Mn and kAB − BAk < δ(ε, n), then A
and B are within ε of a commuting pair. (See [35].) For this reason, the problem
is much less interesting for fixed n. Pearcy and Shields [36] obtain the explicit
estimate δ(ε, n) = 2ε2 /n when A and A0 are Hermitian but B is arbitrary.
10 K.R.Davidson

After these initial results, a variety of counterexamples were found to var-

ious versions of the problem. Voiculescu [43] used very deep methods to estab-
lish the existence of triples An , Bn , Cn of norm one Hermitian n × n matrices
which asymptotically commute but are bounded away from commuting Hermitian
triples. An explicit and somewhat stronger example due to the author [13] pro-
vides matrices An = A∗n and normal matrices Bn in the unit ball of Mn2 +1 with
kAn Bn −Bn An k = n−2 , but bounded away from commuting pairs A0n and Bn0 with
A0n Hermitian but Bn0 are arbitrary. Voiculescu [44] also constructs asymptotically
commuting unitary matrices which are bounded away from commuting unitaries.
Exel and Loring [20] provide a very slick example in which the pairs of unitaries
are actually bounded away from arbitrary commuting pairs. Finally, we mention a
paper of Choi [11] who also found pairs of arbitrary matrices which asymptotically
commute but are bounded away from commuting pairs.
We sketch the Exel–Loring example [20]. Let Un be the cyclic shift on a basis
e1 , . . . , en and let Vn be the diagonal unitary with eigenvalues ω j , 1 ≤ j ≤ n,
where ω = e2πi/n . Then Un Vn Un−1 Vn−1 = ωIn . In particular,
kUn Vn − Vn Un k = |1 − ω| = 2 sin π/n.
Now if An and Bn are commuting matrices within 1/3 of Un and Vn , define As =
(1 − s)Un + sAn and Bs = (1 − s)Vn + sBn . One can check that for 0 ≤ s, t ≤ 1,
γ(s, t) = det (1 − t)In + tAs Bs A−1 −1


s Bs

is never 0; and γ(s, 0) = γ(s, 1) = 1. For fixed s and 0 ≤ t ≤ 1, γ(s, ·) determines

a closed loop in C \ {0}. When s = 0, it reduces to the loop (1 − t + tω)n , which
has winding number 1. But at s = 1, it is the constant loop 1, which has winding
number 0. This establishes a contradiction.
With all this negative evidence, one might suspect that the Hermitian pair
question would also have a negative answer. However, the examples all use some
kind of topological obstruction, which Loring [34] and Loring–Exel [21] make pre-
cise. The case of a pair of Hermitian matrices is different.
This author [13] provided a partial answer to the Hermitian case by proving
an absorption result. If T ∈ Mn is an arbitrary matrix, there is a normal matrix
N in Mn with kN k ≤ kT k and a normal matrix N 0 in M2n so that
kT ⊕ N − N 0 k ≤ 75kT ∗ T − T T ∗ k1/2 .
Thus if T has a small commutator, one obtains a normal matrix close to T ⊕ N .
While this does not answer the question exactly, it can take advantage of an
approximate normal summand of T —and in the case of an essentially normal
operator T , such a summand is available with spectrum equal to σe (T ). The real
problem is that the normal matrix N may have too much spectrum, in some sense.
This approach was pursued in a paper by Berg and the author [7] in order to
provide an operator theoretic proof of the BDF Theorem. The key was to establish
a variant of the absorption theorem for the annulus. Specifically, if T is an invertible
operator with kT k ≤ R and kT −1 k ≤ 1, then there are normal operators N and N 0
 Essentially normal operators 11

satisfying the same bounds (so the spectrum lies in the annulus {z : 1 ≤ |z| ≤ R})
such that
kT ⊕ N − N 0 k ≤ 95kT ∗ T − T T ∗ k1/2 .

We established this by showing that the polar decomposition of T almost com-

mutes, and using the normal summand to provide room for the perturbation.
Combining this with an elementary extraction of approximate eigenvectors allows
one to show that, with T as above, there is a normal operator N with spectrum
in the annulus so that kT − N k ≤ 100kT ∗ T − T T ∗ k1/2 .
The second important step of our proof of BDF is to establish that if T is
essentially normal and ind (T − λI) = 0 for λ 6∈ σe (T ), then T is quasidiagonal.
Since the set of quasidiagonal operators is closed, it suffices to work with a small
perturbation. Now T ∼a T ⊕ N where N is a diagonal normal operator with
σ(N ) = σe (N ) = σe (T ). We fatten up the spectrum of N with a small perturbation
so that it is a nice domain with finitely many smooth holes. An approximation
technique replaces T ⊕ N by a finite direct sum of operators with topologically
annular spectra. Essentially one cuts the spectrum into annular regions without
cutting through any holes. Then the Riesz functional calculus and the case for
the annulus do the job. This results in a proof of the BDF theorem for essentially
normal operators (i.e. planar X).
The almost commuting matrix question was finally solved by Huaxin Lin [33].
This paper is a tour de force. It starts with an idea of Voiculescu’s. Suppose that
there is a counterexample, namely asymptotically commuting n × n Hermitian
matrices An and Bn of norm 1 which are bounded away from commuting pairs.
Let Tn = An + iBn . Then T = T1 ⊕ T2 ⊕ . . . is a block diagonal, essentially normal
operator which is not a block diagonal compact perturbation of a normal operator.
Q
One should consider T as an element of the von Neumann algebra M := Mn
with commutator T ∗ T − T T ∗ lying in the ideal J =
P
Mn of sequences which
converge to 0. The image t = T +J is a normal element of the quotient algebra. Lin
succeeds in proving that t can be approximated by a normal element having finite
spectrum. Then it is an easy matter to lift the spectral projections to projections in
M, and so approximate T by a normal element in M, which yields a good normal
approximation to all but finitely many Tn . Unfortunately this is an extremely
difficult proof.
Lin’s Theorem was made much more accessible by Friis and Rordam [22],
who provide a short, slick and elementary proof. They begin with the same setup.
Observe that by the spectral theorem, every self-adjoint element of any von Neu-
mann algebra can be approximated arbitrarily well be self-adjoint elements with
finite spectrum. This property, called real rank zero (RR0), passes to quotients
like M/J. In Mn , the invertible matrices are dense. If you prove this by modify-
ing the positive part of the polar decomposition, the estimates are independent
of dimension. Thus the argument can be readily extended to show that the in-
vertible elements are dense in M. This property, called topological stable rank one
12 K.R.Davidson

(tsr1), also passes to quotients. Moreover, we observe that a normal element can
perturbed to an invertible normal operator,
Now let t be a normal element of M/J, and fix ε > 0. Cover the spectrum σ(t)
with a grid of lines spaced ε apart horizontally and vertically. Use the tsr1 property
to make a small perturbation which is normal and has a hole in the spectrum inside
each square of the grid. Then use the continuous functional calculus to obtain
another perturbation to a normal element nearby that has spectrum contained in
the grid.
The RR0 property says that self-adjoint elements can be approximated by
self-adjoint elements with finite spectrum. In particular, if a = a∗ has σ(a) = [0, 1],
this property allows us to find b = b∗ with ka − bk < ε such that b − 1/2 is
invertible. A modification of this idea works on each line segment of the grid. So
another small perturbation yields a normal operator with spectrum contained in
the grid minus the mid-point of each line segment. A further use of the functional
calculus collapses the remaining components of the spectum to the lattice points
of the grid. This produces the desired normal approximation with finite spectrum.
In their sequel [23], Friis and Rordam use similar methods in the Calkin
algebra provided that the index data is trivial. They establish quasidiagonality for
essentially normal operators with zero index data, and thus provide a third proof
of the BDF theorem.
Finally we mention that a paper has recently been posted on the arXiv by
Hastings [29] which claims a constructive proof that almost commuting Hermit-
ian matrices are close to commuting, with explicit estimates. This is a welcome
addition since the soft proof provides no norm estimates at all. It is still an open
question whether a perturbation of size O(kT ∗ T − T T ∗ k1/2 ) is possible as in the
case of the absorption results.

References
[1] J. Anderson, A C*-algebra A for which Ext(A) is not a group, Ann. Math. 107
(1978), 455–458.
[2] W. Arveson, A note on essentially normal operators, Proc. Royal Irish Acad. 74
(1974), 143–146.
[3] W. Arveson, Notes in extensions of C*-algebras, Duke Math. J. 44 (1977), 329–355.
[4] M. Atiyah, K-theory, W.A. Benjamin Inc., New York, 1967.
[5] H. Bercovici and D. Voiculescu, The analogue of Kuroda’s theorem for n-tuples, The
Gohberg anniversary collection, Vol. II (Calgary, AB, 1988), 57–60, Oper. Theory
Adv. Appl. 41, Birkhuser, Basel, 1989.
[6] I.D. Berg, An extension of the Weyl-von Neumann theorem to normal operators,
Trans. Amer. Math. Soc. 160 (1971), 365–371.
[7] I.D. Berg and K. Davidson, Almost commuting matrices and a quantitative version
of the Brown-Douglas-Fillmore theorem, Acta Math. 166 (1991), 121–161.
 Essentially normal operators 13

[8] Bratteli, O., Inductive limits of finite dimensional C*-algebras, Trans. Amer. Math.
Soc. 171 (1972), 195–234.
[9] L. Brown, R. Douglas and P. Fillmore, Unitary equivalence modulo the compact op-
erators and extensions of C*-algebras, Proc. conference on Operator theory, Halifax,
NS, Lect. Notes Math. 3445, Springer Verlag, Berlin, 1973.
[10] L. Brown, R. Douglas and P. Fillmore, Extensions of C*-algebras and K-homology,
Ann. Math. 105 (1977), 265–324.
[11] M.D. Choi, Almost commuting matrices need not be nearly commuting, Proc. Amer.
Math. Soc. 102 (1988), 529–533.
[12] M.D. Choi and E. Effros, The completely positive lifting problem for C*-algebras,
Ann. Math. 104 (1976), 585–609.
[13] K. Davidson, Almost commuting Hermitian matrices, Math. Scand. 56 (1985), 222–
240.
[14] K. Davidson, Normal operators are diagonal plus Hilbert-Schmidt, J. Operator The-
ory 20 (1988) 241–250.
[15] K. Davidson, C*-Algebras by Example, Fields Institute Monograph Series 6, Amer-
ican Mathematical Society, Providence, RI, 1996.
[16] R. Douglas, C*-algebra extensions and K-homology, Annals Math. Studies 95, Prince-
ton University Press, Princeton, N.J., 1980.
[17] E. Effros, D. Handelman and C. Shen, Dimension groups and their affine transfor-
mations, Amer. J. Math. 102 (1980), 385–402.
[18] G. Elliott, On the classification of inductive limits of sequences of semi-simple finite
dimensional algebras, J. Algebra 38 (1976), 29–44.
[19] G. Elliott, On totally ordered groups and K0 , Proc. Ring Theory conf., Waterloo, D.
Handelman and J. Lawrence (eds.), Lect. Notes Math. 734 (1978), 1–49, Springer–
Verlag, New York, 1978.
[20] R. Exel and T. Loring, Almost commuting unitary matrices, Proc. Amer. Math. Soc.
106 (1989), 913–915.
[21] R. Exel and T. Loring, Invariants of almost commuting unitaries, J. Funct. Anal. 95
(1991), 364–376.
[22] P. Friis and M. Rordam, Almost commuting self-adjoint matrices—a short proof of
Huaxin’s theorem, J. reine angew. Math. 479 (1996), 121–131.
[23] P. Friis and M. Rordam, Approximation with normal operators with finite spectrum,
and an elementary proof of the Brown–Douglas–Fillmore theorem, Pacific J. Math.
199 (2001), 347–366.
[24] U. Haagerup and S. Thorbjrnsen, A new application of random matrices:
Ext(C∗red (F2 )) is not a group, Ann. Math. 162 (2005), 711–775.
[25] D. Hadwin, An operator valued spectrum, Indiana Univ. Math. J. 26 (1977), 329–340.
[26] P.R. Halmos, Ten problems in Hilbert space, Bull. Amer. Math. Soc. 76 (1970), 887–
933.
[27] P.R. Halmos, What does the spectral theorem say?, Amer. Math. Monthly 70 (1963),
241–247.
[28] P.R. Halmos, Some unsolved problems of unknown depth about operators on Hilbert
space, Proc. Roy. SocĖdinburgh Sect. A 76 (1976/77), 67–76.
14 K.R.Davidson

[29] M. Hastings, Making almost commuting matrices commute, preprint,

arXiv:0808.2474, 2008.
[30] G. Kasparov, The operator K-functor and extensions of C*-algebras (Russian), Izv.
Akad. Nauk SSSR Ser. Mat. 44, (1980), 571–636.
[31] T. Kato, Perturbation of continuous spectra by trace class operators, Proc. Japan
Acad. 33 (1957), 260–264.
[32] S. Kuroda, On a theorem of Weyl–von Neumann, Proc. Japan Acad. 34 (1958),
11–15.
[33] H. Lin, Almost commuting Hermitian matrices and applications, Fields Institute
Comm. 13 (1997), 193–233.
[34] T. Loring, K-theory and asymptotically commuting matrices, Canad. J. Math. 40
(1988), 197216.
[35] W. Luxemburg and R. Taylor, Almost commuting matrices are near commuting ma-
trices, Indag. Math. 32 (1970), 96–98.
[36] C. Pearcy and A. Shields, Almost commuting matrices, J. Funct. Anal. 33 (1979),
332–338.
[37] M. Rosenblum, Perturbation of the continuous spectrum and unitary equivalence,
Pacific J. Math. 7 (1957), 997–1010.
[38] P. Rosenthal, Research Problems: Are Almost Commuting Matrices Near Commuting
Matrices?, Amer. Math. Monthly 76 (1969), 925–926.
[39] W. Sikonia, The von Neumann converse of Weyl’s theorem, Indiana Univ. Math. J.
21 (1971/1972), 121–124.
[40] D. Voiculescu, A non-commutative Weyl–von Neumann Theorem, Rev. Roum. Pures
Appl. 21 (1976), 97–113.
[41] D. Voiculescu, Some results on norm-ideal perturbations of Hilbert space operators,
J. Operator Theory 2 (1979), 3–37.
[42] D. Voiculescu, Some results on norm-ideal perturbations of Hilbert space operators,
II, J. Operator Theory 5 (1981), 77–100.
[43] D. Vioculescu, Remarks on the singular extension in the C*-algebra of the Heisenberg
group, J. Operator Theory 5 (1981), 147–170.
[44] D. Vioculescu, Asymptotically commuting finite rank unitaries without commuting
approximants, Acta Sci. Math. (Szeged) 45 (1983), 429–431.
[45] J. von Neumann, Charakterisierung des Spektrums eines Integraloperators, Actualits
Sci. Indust. 229, Hermann, Paris, 1935.
[46] H. Weyl, Über beschränkte quadratische Formen deren Differ enz vollstetig ist, Rend.
Circ. MatṖalermo 27 (1909), 373–392.

Kenneth R. Davidson
Pure Math. Dept.
U. Waterloo
Waterloo, ON N2L–3G1
CANADA
e-mail: [email protected]