THE GRADUATE STUDENT SECTION

WHAT I S …

a Complex Symmetric
Operator?
Stephan Ramon Garcia
Communicated by Steven J. Miller and Cesar E. Silva

What do these matrices have in common: best explained in the framework of complex symmetric
0 7 0 9 8 9 operators, a surprisingly large class of tractable and
0 1 1 2 ⎡
[ ],[ ] , ⎢0 1 5⎤⎥ , and ⎡
⎢0 7 0⎤ ⎥? well-behaved operators.
0 0 3 4 Let ℋ be a complex Hilbert space. Examples include
⎣0 0 6⎦ ⎣0 0 7⎦
ℂ𝑛 , the Lebesgue spaces 𝐿2 (𝑋, 𝜇) of square-integrable
They each possess a well-hidden symmetry, for they are functions on 𝑋 with respect to a measure 𝜇, the spaces
unitarily similar to the symmetric, but non-Hermitian, ℓ2 (ℕ) and ℓ2 (ℤ) of square integrable sequences indexed
matrices by ℕ and ℤ, and the Hardy Hilbert space 𝐻2 of holomorphic
5−√34
− 12 − 2𝑖 − 2𝑖 functions on the unit disk with square-summable Taylor
[ ], ⎡ 2 ⎤, coefficients at the origin. A conjugate-linear, isometric,
− 2𝑖 1
2 ⎣ −2
𝑖 5+√34
2 ⎦ involution 𝐶 ∶ ℋ → ℋ is a conjugation on ℋ; these are
the Hilbert space analogues of complex conjugation. An
2 + √ 57 0 −1 2
2 𝑖√37 − 73√ 57
⎡
2
⎤ example is [𝐶𝑓](𝑥) = 𝑓(1 − 𝑥) on 𝐿2 [0, 1].
⎢
⎢ 2 ⎥ ⎥
⎢ 0 2 − √ 57
2 2 𝑖√37 + 73√ 57 ⎥ ,
−1 A linear operator 𝑇 ∶ ℋ → ℋ is bounded if ‖𝑇‖ ∶=
sup{‖𝑇x‖ ∶ ‖x‖ ≤ 1} is finite. A bounded linear operator
⎣− 12 𝑖√37 − 73√ 57
2 −1 2
2 𝑖√37 + 73√ 57 3 ⎦
𝑇 ∶ ℋ → ℋ is 𝐶-symmetric if 𝑇 = 𝐶𝑇∗ 𝐶; it is complex
and symmetric if 𝑇 is 𝐶-symmetric with respect to some
8 − √149 9 𝑖 √16837+64√149 𝑖√
133672−1296√149 𝐶. Unbounded examples appear in the complex scaling
2 2
⎡
⎢
√13093 √13093
⎤
⎥ theory for Schrödinger operators, certain non-self-adjoint
⎢
⎢ 9 𝑖 √16837+64√149 207440+9477√149 18√3978002+82324√149 ⎥
⎥, boundary value problems, and 𝒫𝒯-symmetric quantum
2 26186 13093
⎢ √13093 ⎥
133672−1296√149 18√3978002+82324√149 92675+1808√149
theory [1].
⎣𝑖 √ √13093 13093 13093 ⎦ What is the relationship between complex symmetric
respectively. (𝑛 × 𝑛 matrices 𝐴 and 𝐵 are unitarily similar operators and complex symmetric matrices? If 𝐶 is a
if 𝐴 = 𝑈∗ 𝐵𝑈, where 𝑈 is unitary and 𝑈∗ is its adjoint; conjugation on ℋ, then there is an orthonormal basis
operator theorists prefer the term unitarily equivalent (e𝑛 ) of ℋ whose elements are fixed by 𝐶: 𝐶e𝑛 = e𝑛
instead.) The existence of these hidden symmetries is for all 𝑛. Since ⟨𝐶x, 𝐶y⟩ = ⟨y, x⟩ for all x, y ∈ ℋ, the
matrix of a 𝐶-symmetric operator 𝑇 with respect to (e𝑛 )
Stephan Ramon Garcia is associate professor of mathematics at is symmetric:
Pomona College. His e-mail address is Stephan.Garcia@pomona. [𝑇]𝑖,𝑗 = ⟨𝑇e𝑗 , e𝑖 ⟩ = ⟨𝐶𝑇∗ 𝐶e𝑗 , e𝑖 ⟩ = ⟨𝐶e𝑖 , 𝑇∗ 𝐶e𝑗 ⟩
edu.
= ⟨𝑇e𝑖 , e𝑗 ⟩ = [𝑇]𝑗,𝑖 .
The author was partially supported by National Science Founda-
tion grant DMS-1265973. For example, 𝑇 = 𝐶𝑇∗ 𝐶 for
For permission to reprint this article, please contact: 0 1 0 𝑧1 𝑧3
[email protected]. 𝑇=⎡
⎢0 0 1⎤⎥ and 𝐶⎡
⎢𝑧2 ⎤ ⎢𝑧2 ⎤
⎥=⎡ ⎥.
DOI: http://dx.doi.org/10.1090/noti1454 ⎣0 0 0⎦ ⎣𝑧3 ⎦ ⎣𝑧1 ⎦

January 2017 Notices of the AMS 35

 T HE GRA DUATE STUDENT SECTION
Form a unitary early twentieth century in various contexts: T. Takagi
1 1 − 𝑖 (function theory, 1925), N. Jacobson (projective geometry,
2 2 √2
⎡ 1 ⎤ 1939), C.L. Siegel (symplectic geometry, 1943), L.-K. Hua
𝑈=⎢ 0 ⎥
− 1
⎢
⎢ √2 √2
⎥
⎥, (automorphic functions of matrices, 1944), and I. Schur
1 1 𝑖 (quadratic forms, 1945).
⎣ 2 2 √2 ⎦
The innocent-looking Volterra operator
each of whose columns is fixed by 𝐶, and perform the 𝑥
corresponding change of basis: [𝑇𝑓](𝑥) = ∫ 𝑓(𝑦) 𝑑𝑦
0
− 1 0 − 2𝑖
⎡ √2 ⎤ 2
on 𝐿 [0, 1] is a familiar counterexample to many conjec-
𝑈∗ 𝑇𝑈 = ⎢
⎢ 0
⎢ 1 𝑖
2
⎥
⎥
⎥.
√2 tures made by budding operator theorists. For instance,
⎣ − 2𝑖 𝑖
2 0 ⎦ it has no eigenvalues and it is properly quasinilpotent:
Voilá! A hidden symmetry is revealed! There are now ‖𝑇𝑛 ‖1/𝑛 → 0 and 𝑇𝑛 ≠ 0 for 𝑛 = 0, 1, 2, …. It is a standard
procedures to test for the existence of a compatible example of a complex symmetric operator: 𝑇 = 𝐶𝑇∗ 𝐶, in
conjugation; this is how some of the matrices above were which [𝐶𝑓](𝑥) = 𝑓(1 − 𝑥). Each element of the orthonor-
discovered. mal basis (𝑒2𝜋𝑖𝑛𝑥 )𝑛∈ℤ is fixed by 𝐶. With respect to this
This suggests a striking result: each square complex basis, the Volterra operator has the matrix
matrix is similar to a complex symmetric matrix. Here is .
⋱ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ..
the proof: every matrix is similar to its Jordan canonical ⎡⋯ 𝑖 0 0 𝑖
− 6𝜋 0 0 0 ⋯⎤
form, and every Jordan block is unitarily similar to a ⎢ 6𝜋 ⎥
⎢⋯ 𝑖 𝑖 ⋯⎥
⎢ 0 4𝜋 0 − 4𝜋 0 0 0 ⎥
complex symmetric matrix (mimic the example above). ⎢ ⎥
⎢⋯ 0 0 𝑖 𝑖
− 2𝜋 0 0 0 ⋯⎥
Thus, 𝐴 = 𝐴T reveals nothing about the Jordan structure ⎢ 2𝜋 ⎥
⎢ ⎥
of 𝐴. On the other hand, 𝐴 = 𝐴∗ ensures that 𝐴 has an ⎢⋯ 𝑖
− 6𝜋 𝑖
− 4𝜋 𝑖
− 2𝜋 1 𝑖 𝑖 𝑖 ⋯ ⎥,
⎢ 2 2𝜋 4𝜋 6𝜋 ⎥
⎢⋯ 𝑖 𝑖 ⎥
orthonormal basis of eigenvectors and only real eigenval- ⎢ 0 0 0 2𝜋 − 2𝜋 0 0 ⋯⎥
⎢ ⎥
ues. How can this be? It takes 2(1 + 2 + ⋯ + 𝑛) = 𝑛2 + 𝑛 ⎢⋯ 0 0 0 𝑖
0 𝑖
− 4𝜋 0 ⋯⎥
⎢ 4𝜋 ⎥
real parameters to specify an 𝑛 × 𝑛 complex symmetric ⎢⋯ ⎥
⎢ 0 0 0 𝑖
6𝜋 0 0 𝑖
− 6𝜋 ⋯⎥
matrix but only 2(1 + 2 + ⋯ + (𝑛 − 1)) + 𝑛 = 𝑛2 real ⎣. . . ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋱⎦
parameters to specify an 𝑛 × 𝑛 Hermitian matrix, since its
diagonal entries are real. These 𝑛 real degrees of freedom in which the (0, 0) entry has been highlighted. This drives
make all the difference! home the fact that 𝑇 is a rank-one perturbation of a
Although less prevalent than their Hermitian counter- skew-Hermitian operator. One might jest that definite
parts, complex symmetric matrices arise throughout integration is the study of a sparse, infinite complex
mathematics and its applications. For instance, suppose 𝑓 symmetric matrix!
is holomorphic on 𝔻, with 𝑓(0) = 0 and 𝑓′ (0) = 1. Then 𝑓 Examples of complex symmetric operators abound. For
is injective if and only if for any distinct 𝑧1 , 𝑧2 , … , 𝑧𝑛 ∈ ℂ, instance, every idempotent operator, normal operator,
the Grunsky-Goluzin inequality truncated Toeplitz operator, and Hankel matrix is a
|| 𝑛 complex symmetric operator. What sort of properties do
|| ∑ 𝑤𝑗 𝑤𝑘 log ( 𝑧𝑗 𝑧𝑘 𝑓(𝑧𝑗 ) − 𝑓(𝑧𝑘 ) |||
⋅ )| they have?
||𝑗,𝑘=1 𝑓(𝑧𝑗 )𝑓(𝑧𝑘 ) 𝑧𝑗 − 𝑧𝑘 ||
An old result of Godič and Lucenko tells us that each
𝑛 unitary 𝑈 acting on a Hilbert space factors as 𝑈 = 𝐶𝐽, in
1
≤ ∑ 𝑤𝑗 𝑤𝑘 log which 𝐶 and 𝐽 are conjugations. This generalizes the fact
𝑗,𝑘=1 1 − 𝑧𝑗 𝑧𝑘
that a planar rotation is the product of two reflections. A
holds for all w = (𝑤1 , 𝑤2 , … , 𝑤𝑛 ) ∈ ℂ𝑛 . This is a Hermitian- similar result holds for any complex symmetric operator:
symmetric inequality: if 𝑇 is 𝐶-symmetric, then 𝑇 = 𝐶𝐽|𝑇|, in which 𝐽 is a
|⟨𝐴w, w⟩| ≤ ⟨𝐵w, w⟩, conjugation that commutes with the positive operator
|𝑇| = √𝑇∗ 𝑇.
in which 𝐵 = 𝐵 is positive semidefinite and 𝐴 = 𝐴T . In
∗
There are occasional parallels between the Hermitian
applications, complex symmetric matrices have appeared
in the study of thermoelastic waves, quantum reaction and complex-symmetric worlds. This should be surprising
dynamics, vertical cavity surface emitting lasers, electric since many “poorly behaved” operators, like Jordan blocks
power modeling, multicomponent transport, and the and the Volterra operator, are complex symmetric. The
numerical simulation of high-voltage insulators. celebrated Courant minimax principle asserts that if 𝐴
The most familiar result about complex symmetric is an 𝑛 × 𝑛 Hermitian matrix, then the (necessarily real)
matrices is the Autonne-Takagi decomposition: if 𝐴 ∈ eigenvalues 𝜆0 ≥ 𝜆1 ≥ ⋯ ≥ 𝜆𝑛−1 of 𝐴 satisfy
M𝑛 (ℂ) and 𝐴 = 𝐴T , then 𝐴 = 𝑈Σ𝑈T , in which 𝑈 is unitary min max x∗ 𝐴x = 𝜆𝑘 .
and Σ is the diagonal matrix of singular values of 𝐴 codim 𝒱=𝑘 x∈𝒱
‖x‖=1
(the square roots of the eigenvalues of the positive semi-
definite matrix 𝐴∗ 𝐴). It was discovered by Léon Autonne On the other hand, Danciger’s minimax principle ensures
in 1915 and subsequently rediscovered throughout the that if 𝐴 = 𝐴T , then its singular values 𝑠0 ≥ 𝑠1 ≥ ⋯ ≥ 𝑠𝑛−1

36 Notices of the AMS Volume 64, Number 1

 THE GRADUATE STUDENT SECTION
satisfy
⎧𝑠2𝑘 if 0 ≤ 𝑘 < 𝑛2 ,
min max Re xT 𝐴x =
codim 𝒱=𝑘 x∈𝒱 ⎨0 if 𝑛
≤ 𝑘 ≤ 𝑛.
‖x‖=1
⎩ 2

The peculiar singular value “skipping” phenomenon

occurs because of significant cancellation in the complex-
valued expression xT 𝐴x. Naturally, appropriate general-
izations for compact operators exist.
We conclude with a complex-symmetric analogue of
Weyl’s criterion from spectral theory. Let 𝜎(𝐴) denote
the spectrum of a bounded linear operator 𝑇; that is, it
is the set of 𝜆 ∈ ℂ for which 𝑇 − 𝜆𝐼 does not have a
bounded inverse. If 𝑇 = 𝑇∗ , then 𝜆 ∈ 𝜎(𝑇) if and only if
there exist unit vectors x𝑛 such that
lim ‖(𝑇 − 𝜆𝐼)x𝑛 ‖ = 0.
𝑛→∞
The familiar equation 𝑇x = 𝜆x characterizes the eigenval-
ues of 𝑇. A similar result holds in the complex-symmetric
setting. If 𝑇 is 𝐶-symmetric, then |𝜆| ∈ 𝜎(√𝑇∗ 𝑇) if and
only if there are unit vectors x𝑛 so that
lim ‖(𝑇 − 𝜆𝐶)x𝑛 ‖ = 0.
𝑛→∞
In particular, the “antilinear eigenvalue problem” 𝑇x =
|𝜆|𝐶x characterizes the singular values of 𝑇. This can
occasionally be used to obtain information about the
spectrum of |𝑇| without computing 𝑇∗ 𝑇 itself.

References
[1] Stephan Ramon Garcia, Emil Prodan, and Mihai Putinar,
Mathematical and physical aspects of complex symmetric
operators, J. Phys. A 47 (2014), no. 35, 353001, 54 pp.
MR 3254868

Photo Credit
Photo of Stephan Ramon Garcia is courtesy of Pomona
College.

ABOUT THE AUTHOR

Stephan Ramon Garcia got his PhD
at UC Berkeley, worked at UC Santa
Barbara, and now teaches at Pomona
College. He is the author of two books
and over seventy research articles in
operator theory, complex analysis, ma-
trix analysis, number theory, discrete
Stephan Ramon
geometry, and other fields. He has
Garcia
coauthored over two dozen articles
with students.

January 2017 Notices of the AMS 37