Molnar

Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

Communications in Mathematical Physics manuscript No.

(will be inserted by the editor)

Transformations on the set of all n-dimensional

subspaces of a Hilbert space preserving principal


arXiv:math/0011029v1 [math.FA] 4 Nov 2000

angles

Lajos Molnár

Institute of Mathematics and Informatics, University of Debrecen, 4010 Debrecen, P.O.Box


12, Hungary
E-mail: molnarl@math.klte.hu

November 19, 2018

To my wife for her unlimited(?) patience

Abstract: Wigner’s classical theorem on symmetry transformations plays a fun-


damental role in quantum mechanics. It can be formulated, for example, in the
following way: Every bijective transformation on the set L of all 1-dimensional
subspaces of a Hilbert space H which preserves the angle between the elements
of L is induced by either a unitary or an antiunitary operator on H. The aim of
this paper is to extend Wigner’s result from the 1-dimensional case to the case
of n-dimensional subspaces of H with n ∈ N fixed.

1. Introduction and statement of the main result

Let H be a (real or complex) Hilbert space and denote B(H) the algebra of all
bounded linear operators on H. By a projection we mean a self-adjoint idempo-
tent in B(H). For any n ∈ N, Pn (H) denotes the set of all rank-n projections
on H, and P∞ (H) stands for the set of all infinite rank projections. Clearly,
Pn (H) can be identified with the set of all n-dimensional subspaces of H. As it
was mentioned in the abstract, Wigner’s theorem describes the bijective trans-
2 L. Molnár

formations on the set L of all 1-dimensional subspaces of H which preserve the


angle between the elements of L. It seems to be a very natural problem to try to
extend this result from the 1-dimensional case to the case of higher dimensional
subspaces (in our recent papers [11], [12], [13] we have presented several other
generalizations of Wigner’s theorem for different structures). But what about the
angle between two higher dimensional subspaces of H? For our present purposes,
the most adequate concept of angles is that of the so-called principal angles (or
canonical angles, in a different terminology). This concept is a generalization of
the usual notion of angles between 1-dimensional subspaces and reads as follows.
If P, Q are finite dimensional projections, then the principal angles between them
(or, equivalently, between their ranges as subspaces) is defined as the arccos of
the square root of the eigenvalues (counted according multiplicity) of the positive
(self-adjoint) finite rank operator QP Q (see, for example, [1, Exercise VII.1.10]
or [7, Problem 559]). We remark that this concept of angles was motivated by
the classical work [3] of Jordan and it has serious applications in statistics, for
example (see the canonical correlation theory of Hotelling [4], and also see the
introduction of [9]). The system of all principal angles between P and Q is de-
noted by ∠(P, Q). Thus, we have the desired concept of angles between finite
rank projections. But in what follows we would like to extend Wigner’s theorem
also for the case of infinite rank projections. Therefore, we need the concept of
principal angles also between infinite rank projections. Using deep concepts of
operator theory (like scalar-valued spectral measure and multiplicity function)
this could be carried out, but in order to formulate a Wigner-type result we need
only the equality of angles. Hence, we can avoid these complications saying that
for arbitrary projections P, Q, P ′ , Q′ on H we have ∠(P, Q) = ∠(P ′ , Q′ ) if and
only if the positive operators QP Q and Q′ P ′ Q′ are unitarily equivalent. This
obviously generalizes the equality of principal angles between pairs of finite rank
projection.
Preserving principal angles between subspaces of a Hilbert space 3

Keeping in mind the formulation of Wigner’s theorem given in the abstract,


we are now in a position to formulate the main result of the paper which, we
believe, also has physical interpretation.

Main Theorem. Let n ∈ N. Let H be a real or complex Hilbert space with


dim H ≥ n. Suppose that φ : Pn (H) → Pn (H) is a transformation with the
property that

∠(φ(P ), φ(Q)) = ∠(P, Q) (P, Q ∈ Pn (H)).

If n = 1 or n 6= dim H/2, then there exists a linear or conjugate-linear isometry


V on H such that

φ(P ) = V P V ∗ (P ∈ Pn (H)).

If H is infinite dimensional, the transformation φ : P∞ (H) → P∞ (H) satisfies

∠(φ(P ), φ(Q)) = ∠(P, Q) (P, Q ∈ P∞ (H)),

and φ is surjective, then there exists a unitary or antiunitary operator U on H


such that

φ(P ) = U P U ∗ (P ∈ P∞ (H)).

As one can suspect from the formulation of our main result, there is a system
of exceptional cases, namely, when we have dim H = 2n, n > 1. In the next
section we show that in those cases there do exist transformations on Pn (H)
which preserve the principal angles but cannot be written in the form appearing
in our main theorem above.

2. Proof

This section is devoted to the proof of our main theorem. In fact, this will follow
from the statements below.
The idea of the proof can be summarized in a single sentence as follows. We
extend our transformation from Pn (H) to a Jordan homomorphism of the algebra
4 L. Molnár

F (H) of all finite rank operators on H which preserves the rank-1 operators.
Fortunately, those maps turn to have a form and using this we can achieve the
desired conclusion. On the other hand, quite unfortunately, we have to work
hard to carry out all the details of the proof that we are just going to begin.
From now on, let H be a real or complex Hilbert space and let n ∈ N. Since our
statement obviously holds when dim H = n, hence we suppose that dim H > n.
In the sequel, let tr be the usual trace functional on operators. The ideal of
all finite rank operators in B(H) is denoted by F (H). Clearly, every element of
F (H) has a finite trace. We denote by Fs (H) the set of all self-adjoint elements
of F (H).
We begin with two key lemmas. In order to understand why we consider the
property (1) in Lemma 1, we note that if ∠(P, Q) = ∠(P ′ , Q′ ) for some finite
rank projections P, Q, P ′ , Q′ , then, by definition, the positive operators QP Q
and Q′ P ′ Q′ are unitarily equivalent. This implies that tr QP Q = tr Q′ P ′ Q′ .
But, by the properties of the trace, we have tr QP Q = tr P QQ = tr P Q and,
similarly, tr Q′ P ′ Q′ = tr P ′ Q′ . So, if our transformation preserves the principal
angles between projections, then it necessarily preserves the trace of the product
of the projections in question. This justifies the condition (1) in the next lemma.

Lemma 1. Let P be any set of finite rank projections on H. If φ : P → P is a


transformation with the property that

tr φ(P )φ(Q) = tr P Q (P, Q ∈ P), (1)

then φ has a unique real-linear extension Φ onto the real-linear span spanR P of
P. The transformation Φ is injective, preserves the trace and satisfies

tr Φ(A)Φ(B) = tr AB (A, B ∈ spanR P). (2)

Proof. For any finite sets {λi } ⊂ R and {Pi } ⊂ P we define

X X
Φ( λi Pi ) = λi φ(Pi ).
i i
Preserving principal angles between subspaces of a Hilbert space 5

P P
We have to show that Φ is well-defined. If i λi Pi = k µk Qk , where {µk } ⊂ R
and {Qk } ⊂ P are finite subsets, then for any R ∈ P we compute

X X X
tr( λi φ(Pi )φ(R)) = λi tr(φ(Pi )φ(R)) = λi tr(Pi R) =
i i i

X X X
tr( λi Pi R) = tr( µk Qk R) = µk tr(Qk R) =
i k k

X X
µk tr(φ(Qk )φ(R)) = tr( µk φ(Qk )φ(R)).
k k

Therefore, we have

X X
tr(( λi φ(Pi ) − µk φ(Qk ))φ(R)) = 0
i k

for every R ∈ P. By the linearity of the trace functional it follows that we have
similar equality if we replace φ(R) by any finite linear combination of φ(R)’s.
This gives us that

X X X X
tr(( λi φ(Pi ) − µk φ(Qk ))( λi φ(Pi ) − µk φ(Qk ))) = 0.
i k i k
P P
The operator ( i λi φ(Pi ) − k µk φ(Qk ))2 , being the square of a self-adjoint
operator, is positive. Since its trace is zero, we obtain that

X X
( λi φ(Pi ) − µk φ(Qk ))2 = 0
i k

which plainly implies that

X X
λi φ(Pi ) − µk φ(Qk ) = 0.
i k

This shows that Φ is well-defined. The real-linearity of Φ now follows from the
definition. The uniqueness of Φ is also trivial to see. From (1) we immediately
obtain (2). One can introduce an inner product on Fs (H) by the formula

hA, Bi = tr AB (A, B ∈ Fs (H))


6 L. Molnár

(the norm induced by this inner product is called the Hilbert-Schmidt norm).
The equality (2) shows that Φ is an isometry with respect to this norm. Thus,
Φ is injective. It follows from (1) that

tr φ(P ) = tr φ(P )2 = tr P 2 = tr P (P ∈ P)

which clearly implies that

tr Φ(A) = tr A (A ∈ spanR P).

This completes the proof of the lemma. ⊓


In what follows we need the concept of Jordan homomorphisms. If A and


B are algebras, then a linear transformation Ψ : A → B is called a Jordan
homomorphism if it satisfies

Ψ (A2 ) = Ψ (A)2 (A ∈ A),

or, equivalently, if

Ψ (AB + BA) = Ψ (A)Ψ (B) + Ψ (B)Ψ (A) (A, B ∈ A).

Two projections P, Q on H are said to be orthogonal if P Q = QP = 0 (this


means that the ranges of P and Q are orthogonal to each other). In this case we
write P ⊥ Q. We denote P ≤ Q if P Q = QP = P (this means that the range
of P is included in the range of Q). In what follows, we shall use the following
useful notation. If x, y ∈ H, then x ⊗ y stands for the operator defined by

(x ⊗ y)z = hz, yix (z ∈ H).

Lemma 2. Let Φ : Fs (H) → Fs (H) be a real-linear transformation which pre-


serves the rank-1 projections and the orthogonality between them. Then there is
an either linear or conjugate-linear isometry V on H such that

Φ(A) = V AV ∗ (A ∈ Fs (H)).
Preserving principal angles between subspaces of a Hilbert space 7

Proof. Since every finite-rank projection is the finite sum of pairwise orthogonal
rank-1 projections, it is obvious that Φ preserves the finite-rank projections.
It follows from [2, Remark 2.2] and the spectral theorem that Φ is a Jordan
homomorphism (we note that [2, Remark 2.2] is about self-adjoint operators on
finite dimensional complex Hilbert spaces, but the same argument applies for
Fs (H) even if it is infinite dimensional and/or real).
We next prove that Φ can be extended to a Jordan homomorphism of F (H).
To see this, first suppose that H is complex and consider the transformation
Φ̃ : F (H) → F (H) defined by

Φ̃(A + iB) = Φ(A) + iΦ(B) (A, B ∈ Fs (H)).

It is easy to see that Φ is a linear transformation which satisfies Φ̃(T 2 ) = Φ̃(T )2


(T ∈ F (H)). This shows that Φ̃ is a Jordan homomorphism.
If H is real, then the situation is not so simple, but we can apply a deep alge-
braic result of Martindale as follows (cf. the proof of [10, Theorem 3]). Consider
the unitalized algebra F (H) ⊕ RI (of course, we have to add the identity only
when H is infinite dimensional). Defining Φ(I) = I, we can extend Φ to the set
of all symmetric elements of the enlarged algebra in an obvious way. Now we
are in a position to apply the results in [8] on the extendability of Jordan homo-
morphisms defined on the set of symmetric elements of a ring with involution.
To be precise, in [8] Jordan homomorphism means an additive map Ψ which,
besides Ψ (s2 ) = Ψ (s)2 , also satisfies Ψ (sts) = Ψ (s)Ψ (t)Ψ (s). But if the ring in
question is 2-torsion free (in particular, if it is an algebra), this second equality
follows from the first one (see, for example, the proof of [15, 6.3.2 Lemma]). The
statements [8, Theorem 1] in the case when dim H ≥ 3 and [8, Theorem 2] if
dim H = 2 imply that Φ can be uniquely extended to an associative homomor-
phism of F (H) ⊕ RI into itself. To be honest, since the results of Martindale
concern rings and hence linearity does not appear, we could guarantee only the
additivity of the extension of Φ. However, the construction in [8] shows that in
the case of algebras, linear Jordan homomorphisms have linear extensions.
8 L. Molnár

To sum up, in every case we have a Jordan homomorphism of F (H) extending


Φ. In order to simplify the notation, we use the same symbol Φ for the extension
as well.

As F (H) is a locally matrix ring (every finite subset of F (H) can be included
in a subalgebra of F (H) which is isomorphic to a full matrix algebra), it follows
from a classical result of Jacobson and Rickart [6, Theorem 8] that Φ can be
written as Φ = Φ1 + Φ2 , where Φ1 is a homomorphism and Φ2 is an antihomo-
morphism. Let P be a rank-1 projection on H. Since Φ(P ) is also rank-1, we
obtain that one of the idempotents Φ1 (P ), Φ2 (P ) is zero. Since F (H) is a simple
ring, it is easy to see that this implies that either Φ1 or Φ2 is identically zero,
that is, Φ is either a homomorphism or an antihomomorphism of F (H). In what
follows we can assume without loss of generality that Φ is a homomorphism.
Since the kernel of Φ is an ideal in F (H) and F (H) is simple, we obtain that Φ
is injective.

We show that Φ preserves the rank-1 operators. Let A ∈ F (H) be of rank


1. Then there is a rank-1 projection P such that P A = A. We have Φ(A) =
Φ(P A) = Φ(P )Φ(A) which proves that Φ(A) is of rank at most 1. Since Φ is
injective, we obtain that the rank of Φ(A) is exactly 1. From the conditions
of the lemma it follows that φ sends rank-2 projections to rank-2 projections.
Therefore, the range of Φ contains an operator with rank greater than 1. We
now refer to Hou’s work [5] on the form of linear rank preservers on operator
algebras. It follows from the argument leading to [5, Theorem 1.3] that either
there are linear operators T, S on H such that Φ is of the form

Φ(x ⊗ y) = (T x) ⊗ (Sy) (x, y ∈ H)

or there are conjugate-linear operators T ′ , S ′ on H such that Φ is of the form

Φ(x ⊗ y) = (S ′ y) ⊗ (T ′ x) (x, y ∈ H). (3)


Preserving principal angles between subspaces of a Hilbert space 9

Suppose that we have the first possibility. By the multiplicativity of Φ we obtain


that

hu, yiT x ⊗ Sv = hu, yiΦ(x ⊗ v) = Φ(x ⊗ y · u ⊗ v) =


(4)
Φ(x ⊗ y)Φ(u ⊗ v) = hT u, SyiT x ⊗ Sv.

This gives us that hT u, Syi = hu, yi for every u, y ∈ H. On the other hand, since
Φ sends rank-1 projections to rank-1 projections, we obtain that for every unit
vector x ∈ H we have T x = Sx. These imply that T = S is an isometry and
with the notation V = T = S we have

Φ(A) = V AV ∗

for every A ∈ Fs (H).


We show that the possibility (3) cannot occur. In fact, similarly to (4) we
have

hu, yiS ′ v ⊗ T ′ x = hS ′ v, T ′ xiS ′ y ⊗ T ′ u (x, y, u, v ∈ H).

Fixing unit vectors x = y = u in H and considering the operators above at T ′ x,


we find that

S ′ v = hS ′ v, T ′ xihT ′ x, T ′ uiS ′ y

giving us that S ′ is of rank 1. Since Φ sends rank-2 projections to rank-2 pro-


jections, we arrive at a contradiction. This completes the proof of the lemma.

We are now in a position to present a new proof of the nonsurjective version


of Wigner’s theorem which is equivalent to the statement of our main theorem
in the case when n = 1. For another proof see [16].
To begin, observe that if P, Q are finite rank projections such that tr P Q = 0,
then we have tr(P Q)∗ P Q = tr QP Q = tr P QQ = tr P Q = 0 which implies that
(P Q)∗ (P Q) = 0. This gives us that P Q = 0 = QP . Therefore, P is orthogonal
to Q if and only if tr P Q = 0.
10 L. Molnár

Theorem 3. Let φ : P1 (H) → P1 (H) be a transformation with the property that

tr φ(P )φ(Q) = tr P Q (P, Q ∈ P1 (H)). (5)

Then there is an either linear or conjugate-linear isometry V on H such that

φ(P ) = V P V ∗ (P ∈ P1 (H)).

Proof. By the spectral theorem it is obvious that the real linear span of P1 (H)
is Fs (H). Then, by Lemma 1 we see that there is a unique real-linear extension
Φ of φ onto Fs (H) which preserves the rank-1 projections and, by (5), Φ also
preserves the orthogonality between the elements of P1 (H). Lemma 2 applies to
complete the proof. ⊓

As for the cases when n > 1 we need the following lemma. Recall that we
have previously supposed that dim H > n.

Lemma 4. Let 1 < n ∈ N. Then spanR Pn (H) coincides with Fs (H).

Proof. Since the real-linear span of P1 (H) is Fs (H), it is sufficient to show that
every rank-1 projection is a real-linear combination of rank-n projections. To
see this, choose orthonormal vectors e1 , . . . , en+1 in H. Let E = e1 ⊗ e1 + . . . +
en+1 ⊗ en+1 and define

Pk = E − ek ⊗ ek (k = 1, . . . , n + 1).

Clearly, every Pk can be represented by a (n + 1) × (n + 1) diagonal matrix


whose diagonal entries are all 1’s with the exception of the k th one which is 0.
The equation

λ1 P1 + . . . + λn+1 Pn+1 = e1 ⊗ e1

gives rise to a system of linear equations with unknown scalars λ1 , . . . , λn+1 .


The matrix of this system of equations is an (n + 1) × (n + 1) matrix whose
diagonal consists of 0’s and its off-diagonal entries are all 1’s. It is easy to see
that this matrix is nonsingular, and hence e1 ⊗ e1 (and, similarly, every other
Preserving principal angles between subspaces of a Hilbert space 11

ek ⊗ ek ) is a real-linear combination of P1 , . . . , Pn+1 . This completes the proof.



We continue with a technical lemma.

Lemma 5. Let P, Q be projections on H. If QP Q is a projection, then there are


pairwise orthogonal projections R, R′ , R′′ such that P = R + R′ , Q = R + R′′ .
In particular, we obtain that QP Q is a projection if and only if P Q = QP .

Proof. Let R = QP Q. Since R is a projection whose range is contained in the


range of Q, it follows that R′′ = Q − R is a projection which is orthogonal to R.
If x is a unit vector in the range of R, then we have kQP Qxk = 1. Since P Qx
is a vector whose norm is at most 1 and its image under the projection Q has
norm 1, we obtain that P Qx is a unit vector in the range of Q. Similarly, we
obtain that Qx is a unit vector in the range of P and, finally, that x is a unit
vector in the range of Q. Therefore, x belongs to the range of P and Q. Since
x was arbitrary, we can infer that the range of R is included in the range of P .
Thus, we obtain that R′ = P − R is a projection which is orthogonal to R.
Next, using the obvious relations

P R = RP = R, QR = RQ = R

we deduce

(Q − R)(P − R)(Q − R) =

QP Q − QP R − QRQ + QR − RP Q + RP R + RQ − R = (6)

R − R − R + R − R + R + R − R = 0.

Since A∗ A = 0 implies A = 0 for any A ∈ B(H), we obtain from (6) that


R′ R′′ = (P − R)(Q − R) = 0.
The second part of the assertion is now easy to check. ⊓

We next prove the assertion of our main theorem in the case when 1 < n ∈ N
and H is infinite dimensional.
12 L. Molnár

Theorem 6. Suppose 1 < n ∈ N and H is infinite dimensional. If φ : Pn (H) →


Pn (H) is a transformation such that

∠(φ(P ), φ(Q)) = ∠(P, Q) (P, Q ∈ Pn (H)),

then there exists a linear or conjugate-linear isometry V on H such that

φ(P ) = V P V ∗ (P ∈ Pn (H)).

Proof. By Lemma 1 and Lemma 4, φ can be uniquely extended to an injective


real-linear transformation Φ on Fs (H). The main point of the proof is to show
that Φ preserves the rank-1 projections. In order to verify this, just as in the
proof of Lemma 4, we consider orthonormal vectors e1 , . . . , en+1 in H, define
E = e1 ⊗ e1 + . . . + en+1 ⊗ en+1 and set

Pk = E − ek ⊗ ek (k = 1, . . . , n + 1).

We show that the ranges of all Pk′ = φ(Pk )’s can be jointly included in an
(n + 1)-dimensional subspace of H. To see this, we first recall that Φ has the
property that

tr Φ(A)Φ(B) = tr AB (A, B ∈ Fs (H))

(see Lemma 1). Next we have the following property of Φ: if P, Q are orthogonal
rank-1 projections, then Φ(P )Φ(Q) = 0. Indeed, if P, Q are orthogonal, then we
can include them into two orthogonal rank-(n + 1) projections. Now, referring to
the construction given in Lemma 4 and having in mind that Φ preserves the or-
thogonality between rank-n projections, we obtain that Φ(P )Φ(Q) = 0. (Clearly,
the same argument works if dim H ≥ 2(n + 1).) Since the rank-n projections Pk
are commuting, by the preserving property of φ and Lemma 5, it follows that
the projections Φ(Pk ) are also commuting. It is well-known that any finite com-
muting family of operators in Fs (H) can be diagonalized by the same unitary
transformation (or, in the real case, by the same orthogonal transformation).
Therefore, if we resctrict Φ onto the real-linear subspace in Fs (H) generated by
Preserving principal angles between subspaces of a Hilbert space 13

P1 , . . . , Pn+1 , then it can be identified with a real-linear operator from Rn+1


to Rm for some m ∈ N. Clearly, this restriction of Φ can be represented by an
m × (n + 1) real matrix T = (tij ). Let us examine how the properties of Φ are
reflected in those of the matrix T . First, Φ is trace preserving. This gives us that
for every λ ∈ Rn+1 the sums of the coordinates of the vectors T λ and λ are
the same. This easily implies that the sum of the entries of T lying in a fixed
column is always 1. As we have already noted, Φ(ei ⊗ ei )Φ(ej ⊗ ej ) = 0 holds
for every i 6= j. For the matrix T this means that the coordinatewise product of
any two columns of T is zero. Consequently in every row of T there is at most
one nonzero entry. Since Φ sends rank-n projections to rank-n projections, we
see that this possibly nonzero entry is necessarily 1. So, every row contains at
most one 1 and all the other entries in that row are 0’s. Since the sum of the
elements in every column is 1, we have that in every column there is exactly one
1 and all the other entries are 0’s in that column. These now easily imply that if
λ ∈ Rn+1 is such that its coordinates are all 0’s with the exception of one which
is 1, then T λ is of the same kind. What concerns Φ, this means that Φ sends
every ek ⊗ ek (k = 1, . . . , n + 1) to a rank-1 projection.
So, we obtain that Φ preserves the rank-1 projections and the orthogonality
between them. Now, by Lemma 2 we conclude the proof. ⊓

We turn to the case when H is finite dimensional.

Theorem 7. Suppose 1 < n ∈ N, H is finite dimensional and n 6= dim H/2. If


φ : Pn (H) → Pn (H) satisfies

∠(φ(P ), φ(Q)) = ∠(P, Q) (P, Q ∈ Pn (H)),

then there exists a unitary or antiunitary operator U on H such that

φ(P ) = U P U ∗ (P ∈ P∞ (H)). (7)

Proof. First suppose that dim H = 2d, 1 < d ∈ N. If n = 1, . . . , d − 1, then we


can apply the method followed in the proof of Theorem 6 concerning the infinite
14 L. Molnár

dimensional case. If n = d + 1, . . . , 2d − 1, then consider the transformation


ψ : P 7→ I − φ(I − P ) on P2d−n (H). We learn from [7, Problem 559] that if
∠(P, Q) = ∠(P ′ , Q′ ), then there exists a unitary operator U such that U P U ∗ =
P ′ and U QU ∗ = Q′ . It follows from the preserving property of φ that for any
P, Q ∈ P2d−n (H) we have

φ(I − P ) = U (I − P )U ∗ , φ(I − Q) = U (I − Q)U ∗

for some unitary operator U on H. This gives us that

∠(ψ(P ), ψ(Q)) = ∠(U P U ∗ , U QU ∗ ) = ∠(P, Q).

In that way we can reduce the problem to the previous case. So, there is an
either unitary or antiunitary operator U on H such that

ψ(P ) = U P U ∗ (P ∈ P2d−n (H)).

It follows that φ(I − P ) = I − ψ(P ) = I − U P U ∗ = U (I − P )U ∗ , and hence we


have the result for the considered case.
Next suppose that dim H = 2d + 1, d ∈ N. If n = 1, . . . , d − 1, then once
again we can apply the method followed in the proof of Theorem 6. If n =
d + 2, . . . , 2d + 1, then using the ’dual method’ that we have applied right above
we can reduce the problem to the previous case. If n = d, consider a fixed
rank-d projection P0 . Clearly, if P is any rank-d projection orthogonal to P0 ,
then the rank-d projection φ(P ) is orthogonal to φ(P0 ). Therefore, φ induces
a transformation φ0 between d + 1-dimensional spaces (namely, between the
orthogonal complement of the range of P0 and that of the range of φ(P0 )) which
preserves the principal angles between the rank-d projections. Our ’dual method’
and the result concerning 1-dimensional subspaces lead us to the conslusion that
the linear extension of φ0 maps rank-1 projections to rank-1 projections and
preserves the orthogonality between them. This implies that the same holds true
for our original transformation φ. Just as before, using Lemma 1 and Lemma 2
we can conclude the proof. In the remaining case n = d + 1 we apply the ’dual
method’ once again. ⊓

Preserving principal angles between subspaces of a Hilbert space 15

We now show that the case when 1 < n ∈ N, n = dim H/2 is really ex-
ceptional. To see this, consider the transformation φ : P 7→ I − P on Pn (H).
This maps Pn (H) into itself and preserves the principal angles. As for the com-
plex case, the preserving property follows from [1, Exercise VII.1.11] while in
the real case it was proved already by Jordan in [3] (see [14], p. 310). Let us
now suppose that the transformation φ can be written in the form (7). Pick a
rank-1 projection Q on H. We know that it is a real linear combination of some
P1 , . . . , Pn+1 ∈ Pn (H). It would follow from (7) that considering the same linear
combination of φ(P1 ), . . . , φ(Pn+1 ), it is a rank-1 projection as well. But due to
the definition of φ, we get that this linear combination is a constant minus Q. By
the trace preserving property we obtain that this constant is 1/n. Since n > 1,
the operator (1/n)I −Q is obviously not a projection. Therefore, we have arrived
at a contradiction. This shows that the transformation above can not be written
in the form (7).

It would be a nice result if one could prove that in the present case (i.e., when
1 < n, n = dim H/2) up to unitary-antiunitary equivalence, there are exactly
two transformations on Pn (H) preserving principal angles, namely, P 7→ P and
P 7→ I − P . This is left as an open problem.

We now turn to our statement concerning infinite rank projections. In the


proof we shall use the following simple lemma. If A ∈ B(H), then denote by
rng A the range of A.

Lemma 8. Let H be an infinite dimensional Hilbert space. Suppose P, Q are


projections on H with the property that for any projection R with finite corank
we have RP = P R if and only if RQ = QR. Then either P = Q or P = I − Q.

Proof. Let R be any projection on H commuting with P . By Lemma 5, it is


easy to see that we can choose a monotone decreasing net (Rα ) of projections
with finite corank such that (Rα ) converges weakly to R and Rα commutes
with P for every α. Since Rα commutes with Q for every α, we obtain that
16 L. Molnár

R commutes with Q. Interchanging the role of P and Q, we obtain that any


projection commutes with P if and only if it commutes with Q.
Let x be any unit vector from the range of P . Consider R = x ⊗ x. Since R
commutes with P , it must commute with Q as well. By Lemma 5 we obtain that
x belongs either to the range of Q or to its orthogonal complement. It follows
that either d(x, rng Q) = 0, or d(x, rng Q) = 1. Since the set of all unit vectors
in the range of P is connected and the distance function is continuous, we get
that either every unit vector in rng P belongs to rng Q or every unit vector in
rng P belongs to (rng Q)⊥ . Interchanging the role of P and Q, we find that
either rng P = rng Q or rng P = (rng Q)⊥ . This gives us that either P = Q or
P = I − Q. ⊓

Theorem 9. Let H be an infinite dimensional Hilbert space. Suppose that φ :


P∞ (H) → P∞ (H) is a surjective transformation with the property that

∠(φ(P ), φ(Q)) = ∠(P, Q) (P, Q ∈ P∞ (H)).

Then there exists a unitary or antiunitary operator U on H such that

φ(P ) = U P U ∗ (P ∈ P∞ (H)).

Proof. We first prove that φ is injective. If P, P ′ ∈ P∞ (H) and φ(P ) = φ(P ′ ),


then by the preserving property of φ we have

∠(P, Q) = ∠(P ′ , Q) (Q ∈ P∞ (H)). (8)

Putting Q = I, we see that P is unitarily equivalent to P ′ . We distiguish two


cases. First, let P be of infinite corank. By (8), we deduce that for every Q ∈
P∞ (H) we have Q ⊥ P if and only if Q ⊥ P ′ . This gives us that P = P ′ . As
the second possibility, let P be of finite corank. Then P, P ′ can be written in
the form P = I − P0 and P ′ = I − P0′ , where, by the equivalence of P, P ′ , the
projections P0 and P0′ have finite and equal rank. Let Q0 be any finite rank
projection on H. It follows from

∠(I − P0 , I − Q0 ) = ∠(I − P0′ , I − Q0 )


Preserving principal angles between subspaces of a Hilbert space 17

that there is a unitary operator W on H such that

W (I − Q0 )(I − P0 )(I − Q0 )W ∗ = (I − Q0 )(I − P0′ )(I − Q0 ).

This implies that

W (−Q0 − P0 + P0 Q0 + Q0 P0 − Q0 P0 Q0 )W ∗ =

−Q0 − P0′ + P0′ Q0 + Q0 P0′ − Q0 P0′ Q0 .

Taking traces, by the equality of the rank of P0 and P0′ , we obtain that

tr P0 Q0 = tr P0′ Q0 . (9)

Since this holds for every finite rank projection Q0 on H, it follows that P0 = P0′
and hence we have P = P ′ . This proves the injectivity of φ.
Let P ∈ P∞ (H) be of infinite corank. Then there is a projection Q ∈ P∞ (H)
such that Q ⊥ P . By the preserving property of φ, this implies that φ(Q) ⊥ φ(P )
which means that φ(P ) is of infinite corank. One can similarly prove that if φ(P )
is of infinite corank, then the same must hold for P . This yields that P ∈ P∞ (H)
is of finite corank if and only if so is φ(P ).
Denote by Pf (H) the set of all finite rank projections on H. It follows that
the transformation ψ : Pf (H) → Pf (H) defined by

ψ(P ) = I − φ(I − P ) (P ∈ Pf (H))

is well-defined and bijective. Since φ(I − P ) is unitarily equivalent to I − P for


every P ∈ Pf (H) (this is because ∠(φ(I − P ), φ(I − P )) = ∠(I − P, I − P )), it
follows that ψ is rank preserving.
We next show that

tr ψ(P )ψ(Q) = tr P Q (P, Q ∈ Pf (H)). (10)

This can be done following the argument leading to (9). In fact, by the preserving
property of φ there is a unitary operator W on H such that

W (I − ψ(Q))(I − ψ(P ))(I − ψ(Q))W ∗ = (I − Q)(I − P )(I − Q).


18 L. Molnár

This gives us that

W (−ψ(Q) − ψ(P ) + ψ(P )ψ(Q) + ψ(Q)ψ(P ) − ψ(Q)ψ(P )ψ(Q))W ∗ =

−Q − P + P Q + QP − QP Q.

Taking traces on both sides and referring to the rank preserving property of
ψ, we obtain (10). According to Lemma 1, let Ψ : Fs (H) → Fs (H) denote the
unique real-linear extension of ψ onto spanR Pf (H) = Fs (H). We know that Ψ
is injective. Since Pf (H) is in the range of Ψ , we obtain that Ψ is surjective as
well. It is easy to see that Lemma 2 can be applied and we infer that there exists
an either unitary or antiunitary operator U on H such that

Ψ (A) = U AU ∗ (A ∈ Fs (H)).

Therefore, we have

φ(P ) = U P U ∗

for every projection P ∈ P∞ (H) with finite corank. It remains to prove that the
same holds true for every P ∈ P∞ (H) with infinite corank as well. This could be
quite easy to show if we know that φ preserves the order between the elements
of P∞ (H). But this property is far away from being easy to verify. So we choose
a different approach to attack the problem.
Let P ∈ P∞ (H) be a projection of infinite corank. By the preserving property
of φ we see that for every Q ∈ P∞ (H) the operator φ(Q)φ(P )φ(Q) is a projection
if and if QP Q is a projection. By Lemma 5, this means that φ(Q) commutes
with φ(P ) if and only if Q commutes with P . Therefore, for any Q ∈ P∞ (H) of
finite corank, we obtain that Q commutes with U ∗ φ(P )U (this is equivalent to
that φ(Q) = U QU ∗ commutes with φ(P )) if and only if Q commutes with P .
By Lemma 8 we have two possibilities, namely, either U ∗ φ(P )U = P or
U ∗ φ(P )U = I − P . Suppose that U ∗ φ(P )U = I − P . Consider a complete
orthonormal basis e0 , eγ (γ ∈ Γ ) in the range of P and, similarly, choose a
Preserving principal angles between subspaces of a Hilbert space 19

complete orthonormal basis f0 , fδ (δ ∈ ∆) in the range of I − P . Pick nonzero


scalars λ, µ with the property that |λ|2 + |µ|2 = 1 and |λ| 6= |µ|. Define
X X
Q = (λe0 + µf0 ) ⊗ (λe0 + µf0 ) + eγ ⊗ eγ + fδ ⊗ fδ .
γ δ

Clearly, Q is of finite corank (in fact, its corank is 1). Since φ(Q)φ(P )φ(Q) =
U QU ∗ φ(P )U QU ∗ is unitarily equivalent to QP Q, it follows that the spectrum of
QU ∗ φ(P )U Q is equal to the spectrum of QP Q. This gives us that the spectrum
of Q(I − P )Q is equal to the spectrum of QP Q. By the construction of Q this
means that

{0, 1, |µ|2} = {0, 1, |λ|2 }

which is an obvious contradiction. Consequently, we have U ∗ φ(P )U = P , that


is, φ(P ) = U P U ∗ . Thus, we have proved that this latter equality holds for every
P ∈ P∞ (H) and the proof is complete. ⊓

Acknowledgements. This research was supported from the following sources: (1) Hungarian
National Foundation for Scientific Research (OTKA), Grant No. T030082, T031995, (2) A
grant from the Ministry of Education, Hungary, Reg. No. FKFP 0349/2000.

References

1. Bhatia, R.: Matrix Analysis. Springer-Verlag, 1997


2. Brešar, M., Šemrl, P.: Mappings which preserve idempotents, local automorpisms, and
local derivations. Canad. J. Math. 45, 483–496 (1993)
3. Jordan, C.: Essai sur la géométrie á n dimensions. Bull. Soc. Math. France 3, 103–174
(1875)
4. Hotelling, H.: Relations between two set of variates. Biometrika 28, 321-377 (1935)
5. Hou, J.C.: Rank-preserving linear maps on B(X). Sci. China Ser. A 32, 929–940 (1989)
6. Jacobson, N., Rickart, C.: Jordan homomorphisms of rings. Trans. Amer. Math. Soc.
69, 479–502 (1950)
7. Kirillov, A.A., Gvishiani, A.D.: Theorems and Problems in Functional Analysis.
Springer-Verlag, 1982
8. Martindale, W.S.: Jordan homomorphisms of the symmetric elements of a ring with
involution. J. Algebra 5, 232–249 (1967)
20 L. Molnár

9. Miao, J., Ben-Israel, A.: On principal angles between subspaces in Rn . Linear Algebra
Appl. 171 81–98 (1992)
10. Molnár, L.: An algebraic approach to Wigner’s unitary-antiunitary theorem. J. Austral.
Math. Soc. 65, 354–369 (1998)
11. Molnár, L.: A generalization of Wigner’s unitary-antiunitary theorem to Hilbert mod-
ules. J. Math. Phys. 40, 5544–5554 (1999)
12. Molnár, L.: Generalization of Wigner’s unitary-antiunitary theorem for indefinite inner
product spaces. Commun. Math. Phys. 201, 785–791 (2000)
13. Molnár, L.: A Wigner-type theorem on symmetry transformations in type II factors.
Int. J. Theor. Phys. to appear
14. Paige, C.C., Wei, M.: History and generality of the CS decomposition. Linear Algebra
Appl. 208/209, 303-326 (1994)
15. Palmer, T.W.: Banach Algebras and The General Theory of *-Algebras, Vol. I. Cam-
bridge University Press, 1994
16. Sharma, C.S., Almeida, D.F.: A direct proof of Wigner’s theorem on maps which
preserve transition probabilities between pure states of quantum systems. Ann. Phys.
197, 300–309 (1990)

Communicated by H. Araki

You might also like