arXiv:0709.2444v1 [math.RT] 15 Sep 2007
Unitary and Euclidean Representations of a
Quiver
Vladimir V. Sergeichuk
Institute of Mathematics, Tereshchenkivska 3, Kiev, Ukraine
sergeich@imath.kiev.ua
Abstract
A unitary (Euclidean) representation of a quiver is given by assigning to each vertex a unitary (Euclidean) vector space and to each
arrow a linear mapping of the corresponding vector spaces. We recall an algorithm for reducing the matrices of a unitary representation
to canonical form, give a certain description of the representations
of canonical form, and reduce the problem of classifying Euclidean
representations to the problem of classifying unitary representations.
We also describe the set of dimensions of all indecomposable unitary
(Euclidean) representations of a quiver and establish the number of
parameters in an indecomposable unitary representation of a given
dimension.
1
Introduction
Many problems of linear algebra can be formulated and studied in terms of
quivers and their representations, which were proposed by Gabriel [1] (see
also [2]). A quiver is a directed graph. Its representation A is given by
assigning to each vertex i a vector space Ai and to each arrow α : i → j a
linear mapping Aα : Ai → Aj . For example, the canonical form problems
✄
q ♣ correspond to the canonical
for representations of the quivers ✯
✂ ♣ and ♣ ✶
This is the author’s version of a work that was published in Linear Algebra Appl. 278
(1998) 37–62. Partially supported by Grant No. U6E000 from the International Science
Foundation.
1
form problems for linear operators (whose solution is the Jordan normal
form) and for pairs of linear mappings from one space to another (the matrix
pencil problem, solved by Kronecker).
In this chapter we study unitary and Euclidean representations of a quiver
up to isometry. A unitary (Euclidean) representation A is given by assigning
to each vertex i a finite dimensional unitary (Euclidean) space Ai and to
each arrow α : i → j a linear mapping Aα : Ai → Aj . We say that two
unitary (Euclidean) representations A and B are isometric and write A ≃ B
if there exists a system of isometries Φi : Ai → Bi such that Φj Aα = Bα Φi
for each α : i → j.
Our main tool is Littlewood’s algorithm [3] for reducing matrices to triangular canonical form via unitary similarity. In [4] I rediscovered Littlewood’s algorithm and applied it to the canonical form problem for unitary
representations of a quiver. Various algorithms for reducing matrices to different canonical forms under unitary similarity were also proposed by Brenner, Mitchell, McRae, Radjavi, Benedetti and Gragnolini, and others; see
Shapiro’s survey [5].
In Section 2 we recall briefly Littlewood’s algorithm and study the structure of canonical matrices much as it was made in [4] for the matrices of
linear operators in a unitary space.
We say that a matrix problem is unitarily wild if it contains the problem
of classifying linear operators in a unitary space. In Section 2.3 we show that
the last problem contains the problem of classifying unitary representations
of an arbitrary quiver (i.e., it is hopeless in a certain sense) and give examples
of unitarily wild matrix problems.
The vector
dim A = (dimA1 , dimA2 , . . . , dimAp ) ∈ Np0
is called the dimension of a representation A of a quiver Q with vertices
1, 2, . . . , p (we denote
N = {1, 2, . . . },
N0 = {0, 1, 2, . . . }).
In Section 3 we describe the set of dimensions of direct-sum-indecomposable
unitary representations of a quiver, and establish the number of parameters
in an indecomposable unitary representation of a given dimension. Analogous, but much more fundamental and complicated, results for non-unitary
2
representations of a quiver were obtained by V. G. Kac [6, 7, 8] (see also [2,
Sect 7.4]).
In particular, if z ∈ Np and Q is a connected quiver other than • and • → •,
then there exists an indecomposable unitary representation of dimension z if
and only if zMQ ≥ z, where MQ = [mij ] is the p × p matrix whose entry mij
is the number of arrows of the form i → j and i ← j, where
(t1 , . . . , tp ) ≥ (z1 , . . . , zp )
means
t1 ≥ z1 , . . . , tp ≥ zp .
In Section 4 we study Euclidean representations of a quiver. Let AC
denote the unitary representation obtained from a Euclidean representation
A by complexification (A and AC are given by the same set of real matrices).
In Section 4.1 we prove intuitively obvious facts that
(i) AC ≃ BC implies A ≃ B, and
(ii) if A is indecomposable and AC is decomposable, then AC ≃ U ⊕ Ū ,
where U is an indecomposable unitary representation.
This will imply that unitary and Euclidean representations have the same
sets of dimensions of indecomposable representations.
In Section 4.2 we study, when a given unitary representation of a quiver
can be obtained by complexification. In particular, let A be a complex matrix
that is not unitarily similar to a direct sum of matrices, and let S −1 AS = Ā
for a unitary matrix S (such S exists if A is unitarily similar to a real matrix).
Then A is unitarily similar to a real matrix if and only if S is symmetric.
2
Unitary matrix problems
We suppose that the complex numbers are lexicographically ordered:
a + bi a′ + b′ i if either a = a′ and b ≤ b′ , or a < a′ ;
(1)
and that the set of blocks of a block matrix A = [Aij ] are linearly ordered:
Aij ≤ Ai′ j ′ if either i = i′ and j ≤ j ′ , or i > i′ .
3
(2)
A block complex matrix with a given (perhaps empty) set of marked
square blocks will be called a marked block matrix ; a square block is marked
by a line along its principal diagonal. By a unitary matrix problem we mean
the classification problem for marked block matrices
A = [Aij ],
1 ≤ i ≤ l,
1 ≤ j ≤ r,
up to transformations
A 7→ B := R−1 AS = [Ri−1 Aij Sj ],
(3)
where
R = R1 ⊕ · · · ⊕ Rl ,
S = S1 ⊕ · · · ⊕ Sr
are unitary matrices, and Ri = Sj whenever the block Aij is marked. The
transformation (3) is called an admissible transformation; we say that these
marked block matrices A and B (with the same disposition of marked blocks)
are equivalent and write A ∼ B or
(R, S) : A ❀ B.
(4)
Notice that a matrix consisting of a single block is reduced by transformations of unitary similarity if the block is marked, and by transformations of
unitary equivalence otherwise. Moreover, the matrices of every unitary representation A of a quiver can be placed into a marked block matrix A such
that the admissible transformations with A correspond to reselections of the
orthogonal bases in the spaces of A, for example,
2
µ
✞
✠
λ ✝ ✿1 ✛
2.1
ξ
ν
❘
3
S1
S2
S3
S1−1
Aλ
Aµ
Aν
S3−1
0
(5)
Aξ
An algorithm
The algorithm is based on the following two lemmas:
4
0
Lemma 2.1. (a) Each complex matrix A is unitarily equivalent to the matrix
D = a1 I ⊕ · · · ⊕ ak−1 I ⊕ 0,
ai ∈ R,
a1 > · · · > ak−1 > 0.
(6)
(b) If R−1 DS = D ′ , where R and S are unitary matrices and D, D ′ are
of the form (6), then D = D ′ ,
S = S1 ⊕ · · · ⊕ Sk−1 ⊕ S ′ ,
and
R = S1 ⊕ · · · ⊕ Sk−1 ⊕ R′ ,
where each Si has the same size as ai I.
Lemma 2.2. (a) Each square complex matrix A is unitarily similar to the
block-triangular matrix
λ1 · · · λk
λ1 I F12 · · · F1k
(see (1)), the columns
λ2 I · · · F2k
F =
(7)
.. , of Fi,i+1 are linearly
..
.
.
independent if λi =
0
λk I
λ .
i+1
(b) If S −1 F S = F ′ , where S is a unitary matrix and F and F ′ have the
form (7), then λi I = λ′i I and
S = S1 ⊕ · · · ⊕ Sk ,
where each Si has the same size as λi I.
Proof. These lemmas were proved in many articles, see, for example, [3, 4, 5],
so we give only an outline of their proofs. Part (a) of Lemma 2.1 is the
singular value decomposition; part (b) follows from
D = D′,
S ∗ D ∗ R∗−1 = D ∗ ,
D 2 R = RD 2 ,
S −1 DR = D,
D 2 S = SD 2 .
The matrix (7) is the matrix of an arbitrary linear operator A : Cn → Cn in
an orthogonal basis f1 , . . . , fn such that f1 , . . . , ftr is a basis of
Ker(A − λ1 I) · · · (A − λr I),
1 ≤ r ≤ k,
where
(x − λ1 ) · · · (x − λk ),
λ1 · · · λk ,
is the minimal polynomial of A; it proves part (a) of Lemma 2.2. Successively
equating the blocks of F S = SF ′ ordered with (2), we prove part (b).
5
By the canonical part of the matrix (6) or (7), we mean the matrix (6)
or, respectively, the collection of blocks Fij , i ≥ j. According to Lemmas 2.1
and 2.2, the canonical part is uniquely determined by the initial matrix A
and does not change if A is replaced by a unitarily equivalent or, respectively,
similar matrix.
The algorithm for reducing a marked block matrix A = [Aij ] to canonical
form:
Let Apq be the first (in the ordering (2)) block of A that changes under admissible transformations (3). Depending on the arrangement of the
marked blocks, it is reduced by the transformations of unitary equivalence
or similarity. Respectively, we reduce A = [Aij ] to the matrix à = [Ãij ] with
Ãpq of the form (6) or (7), and then restrict ourselves to those admissible
transformations with à that preserve the canonical part of Ãpq . As follows
from Lemmas 2.1(b) and 2.2(b), they are exactly the admissible transformations with the marked block matrix A′ that is obtained in the following
way: The block Ãpq of the form (6) or (7) consists of k horizontal and k
vertical strips; we extend this partition to the whole p-th horizontal and the
whole q-th vertical strips of Ã. If new k divisions pass through the marked
block Ãij , we carry out k perpendicular divisions such that Ãij is partitioned
into k × k subblocks with square diagonal blocks (they are crossed by the
marking line) and repeat this for all new divisions. We additionally mark the
subblocks a1 I, . . . , ak−1 I of Ãpq if it has the form (6). The obtained marked
block matrix A′ will be called the derived matrix of A. Clearly, A ∼ B implies
A′ ∼ B ′ .
Let us consider the sequence of derived matrices
A(0) := A, A′ , A′′ , . . . , A(s) .
(8)
This sequence ends with a certain matrix A(s) , s ≥ 0, for which the admissible
transformations do not change any of its blocks, i.e, A(s) is equivalent only
to itself. Then A ∼ B implies A(s) ∼ B (s) , i.e., A(s) = B (s) . Remove from
A(s) all additional divisions into subblocks and additional marking lines that
have appeared during the reduction of A to A(s) . The obtained marked block
matrix will be called a canonical matrix or the canonical form of A and will
be denoted by A∞ . We have the following
Theorem 2.1. Each marked block matrix A is equivalent to the uniquely
determined canonical matrix A∞ ; moreover, A ∼ B if and only if A∞ = B ∞ .
✷
6
We will take under consideration the null matrices 00n , 0m0 , and 000 of
size 0 × n, m × 0, and 0 × 0, putting for a p × q matrix M
M
,
M ⊕ 00n = M 0pn , M ⊕ 0m0 =
0mq
0m0 ⊕ 00n = 0mn . Respectively, we will consider block matrices with “empty”
horizontal and/or vertical strips.
Let A = [Aij ] and B = [Bij ] (1 ≤ i ≤ l, 1 ≤ j ≤ r) be marked block
matrices with the same set of indices (i, j) of the marked blocks. By the block
direct sum of A and B we mean the marked block matrix
A ⊎ B := [Aij ⊕ Bij ]
with the same disposition of marked blocks. If
T1 = (R1 , S1 ) : A ❀ C
and
T2 = (R2 , S2 ) : B ❀ D
(see (4)), then R1 , R2 and, respectively, S1 , S2 are block diagonal matrices
with l and, respectively, r diagonal square blocks, and
T1 ⊎ T2 := (R1 ⊎ R2 , S1 ⊎ S2 ) : A ⊎ B ❀ C ⊎ D.
A marked block matrix A is said to be indecomposable if
(i) its size other than 0 × 0, and
(ii) A ∼ B ⊎ C implies that B or C has size 0 × 0.
For every matrices M1 , . . . , Mn , N, we define
(M1 , . . . , Mn ) ⊗ N := (M1 ⊗ N, . . . , Mn ⊗ N),
(9)
where Mi ⊗ N is obtained from Mi by replacing its entries a with aN.
Theorem 2.2. (a) Each marked block matrix A is equivalent to a matrix of
the form
B = (P1 ⊗ Im1 ) ⊎ · · · ⊎ (Pt ⊗ Imt )
∼ P1 ⊎ · · · ⊎ P1 ⊎ · · · ⊎ Pt ⊎ · · · ⊎ Pt ,
|
|
{z
}
{z
}
m1 copies
mt copies
7
where P1 , . . . , Pt are nonequivalent indecomposable marked block matrices,
uniquely determined up to equivalence (we may take P1 = P1∞ , . . . , Pt = Pt∞ ),
and m1 , . . . , mt are uniquely determined natural numbers. Every admissible
transformation T : B ❀ B that preserves B has the form
T = (1P1 ⊗ U1 ) ⊎ · · · ⊎ (1Pt ⊗ Ut ),
where 1Pi = (I, I) : Pi ❀ Pi is the identity transformation of Pi , and Ui is a
unitary mi × mi matrix (1 ≤ i ≤ t).
(b) A marked block matrix A of size 6= 0 × 0 is indecomposable if and
only if every preserving it admissible transformation T : A ❀ A has the form
T = a1A , a ∈ C, |a| = 1.
(c) A canonical matrix can be reduced to an equivalent block direct sum
of indecomposable canonical matrices using only admissible permutations of
rows and columns.
Proof. (a) We may take A = A∞ . Since admissible transformations with
A(i) , 1 ≤ i ≤ s, (see (8)) are exactly the admissible transformations with A
that preserve the already reduced part of A(i) (preserve A(s) if i = s), the set
of admissible transformations with A(s) consists of all (R, S) : A ❀ A. By
(3),
R = R1 ⊕ · · · ⊕ Rl ,
S = S1 ⊕ · · · ⊕ Sr ,
where l × r is the number of blocks of A. Since (R, S) : A(s) ❀ A(s) , we have
Ri = Uf (i,1) ⊕ · · · ⊕ Uf (i,li ) ,
Sj = Ug(j,1) ⊕ · · · ⊕ Ug(j,rj ) ,
(10)
where
f (i, α), g(j, β) ∈ {1, . . . , t}
and U1 , . . . , Ut are arbitrary unitary matrices of fixed sizes. A(s) differs from
A only by additional divisions of its strips into substrips (and by additional
marking lines). We transpose substrips within each strip of A to obtain a
matrix B ∼ A such that, for all (R, S) : B ❀ B, we have (10) with
f (i, 1) ≤ · · · ≤ f (i, li),
g(j, 1) ≤ · · · ≤ g(j, rj ).
Clearly, B satisfies (a).
(b)&(c) These statemants are obvious.
8
2.2
The structure of canonical matrices
In this section we divide the set of canonical m × n matrices into disjoint
subsets of canonical matrices with the same “scheme” (the number of such
schemes is finite for each size m × n), and show how to construct all the
canonical matrices with a given scheme (for matrices under unitary similarity
this was made briefly in [4]).
We partition a canonical matrix into zones, which illustrate the reduction
process.
Let A = A∞ be a canonical matrix. Then all its derived matrices (8) differ
from A only by additional divisions and marking lines. Denote by Pl (0 ≤
l < s) the first block of A(l) that changes under admissible transformations
(it is reduced when we construct A(l+1) ).
(l)
(l)
Let Aij be a block of A(l) such that either Aij ≤ Pl or l = s. The admissible transformations with A(l) induce the unitary equivalence or similarity
(l)
(l)
transformations with Aij . Respectively, Aij has the form (6) or (7); we de(l)
note by Z(Aij ) its canonical part (see page 6). Defining by induction in l,
(l)
(l)
we call Z(Aij ) by a zone and l by its depth if either l = 0 or Z(Aij ) is not
contained in a zone of depth < l.
(l)
(l)
For each zone Z = Z(Aij ), we put Bl(Z) := Aij and call Z by an
equivalence (similarity) zone if Bl(Z) is transformed by unitary equivalence
(similarity) transformations.
Clearly, every canonical matrix A is partitioned into equivalence and similarity zones; for example (for a marked block matrix of the form ❅❅
),
9
A=
i 0 0 0 2 0 0 0 3
❅
0 ❅i 0 0 0 2 0 0 0
0 0❅i 0 0 0 2 0 0
0 0 0❅i❅0 0 0 0 0
0 0 0 0 0❅ 0 0 0 5
0 0 0 0 0❅
0 0 0 0
0 0 0 0 0 0❅
0 0 0
0 0 0 0 0 0 0❅
0 0
❅
0
3
0
0
0
5
0
0
2
0
i
0
0
0
5
0
i
0
4
4
0
0
0
3
7 8
8
5
6
4 4
3
=
1
(11)
2
is partitioned into 10 zones, their depths are indicated on the right of (11).
Let A be a canonical matrix partitioned into zones. For each similarity
zone, we replace all its diagonal elements by stars. For each equivalence zone,
we replace all its nonzero elements by circles, and join with a line its circles
corresponding to equal elements (this line does not coincide with a marking
line because the marking lines connect stars). The other elements of A are
zeros, we replace theirs by points. The obtained picture will be called the
scheme S(A) of A.
For example, the canonical matrix (11) has the scheme
S(A) =
⋆
·
·
·
·
·
·
·
·
⋆
·
·
·
·
·
·
·
·
⋆
·
·
·
·
·
·
·
·
⋆
·
·
·
·
•
·
·
·
⋆
·
·
·
·
•
·
·
·
⋆
·
·
·
·
•
·
·
·
⋆
·
·
·
·
·
·
·
·
⋆
⋆
·
·
·
•
·
·
·
·
⋆
·
·
·
•
·
·
•
·
⋆
·
·
·
•
·
⋆
·
•
•
·
·
·
•
Theorem 2.3. Each canonical matrix A = [aij ] with a given scheme S = [sij ]
can be constructed by successive filling of its zones by numbers starting with
10
the zones of greatest depth as follows: Let Z be a zone of depth d(Z) and
let all entries in zones of depth > d(Z) be replaced by numbers. Then we
replace all points, circles, and stars of Z, respectively, by zeros, positive real
numbers, and complex numbers such that the following conditions hold:
1) Let sij and si+1,j+1 be circles in Z. Then aij = ai+1,j+1 if sij and
si+1,j+1 are linked by a line, and aij > ai+1,j+1 otherwise.
2) Let sα,β , . . . , sα+k,β+k be all stars of Z that lie under a certain stair of
Z. Then
aα,β = · · · = aα+k,β+k .
If
sα+k+1,β+k+1, . . . , sα+t,β+t
are all stars of Z that lie under the next stair of Z, then aα,β aα+t,β+t ;
moreover, aα,β ≻ aα+t,β+t whenever the columns of the block
[ aij | α ≤ i ≤ α + k, β + k + 1 ≤ j ≤ β + t]
are linearly dependent (this block has been filled by numbers because all its
entries are located in zones of depth > d(Z)).
✷
This theorem gives a convenient way to present solutions of unitary matrix
problems in small sizes by their sets of schemes. Thus, the list of schemes of
canonical 5 × 5 matrices under unitary similarity was obtained by Klimenko
[11].
2.3
Unitarily wild matrix problems
The canonical form problem for pairs of n × n matrices under simultaneous
✄
similarity (i.e., for representations of the quiver ✯
✂ ♣ ❨✁ ) plays a special role
in the theory of (non-unitary) matrix problems. It may be proved that its
solution implies the clasification of representations of every quiver (and even
representations of every finite dimensional algebra). For this reason, the
classification problem for pairs of matrices under simultaneous similarity is
used as a yardstick of the complexity; Donovan and Freislich [12] (see also [2])
suggested to name a classification problem wild if it contains the problem of
simultaneous similarity, and otherwise to name it tame (in accordance with
the partition of animals into wild and tame ones).
The canonical form problem for an n × n matrix under unitary similarity
✄ ♣ ) plays the same role in the
(i.e., for unitary representations of the quiver ✯
✂
11
theory of unitary matrix problems: it contains the problem of classifying unitary representations of every quiver. For example, the problem of classifying
unitary representations of the quiver (5) can be regarded (by Lemma 2.2) as
the problem of classifying, up to unitary similarity, matrices of the form:
5I I Aλ Aν Aµ
0 4I I
0
0
0
.
0
3I
0
0
0
0
0 2I Aξ
0
0
0
0
I
A matrix problem is called unitarily wild (or *-wild, see [13]) if it contains
the problem of classifying matrices via unitary similarity, and unitarily tame
otherwise.
For each unitary problem, one has an alternative: to solve it or to prove
that it is unitarily wild (and hence is hopeless in a certain sense). In this
section we give some examples of such alternatives.
(i) Let us consider the problems of classifying nilpotent linear operators
ϕ, ϕn = 0, in a unitary space.
For n = 2 this problem is unitarily tame; the canonical matrix of ϕ (see
page 6) is
0 D
,
0 0
where D is of the form (6) without zero columns. Indeed, a matrix F = [Fij ]
of the form (7) satisfies F 2 = 0 only if k = 2, F11 = 0, and F22 = 0; we can
reduce F12 to the form (6).
For n > 2 this problem is unitarily wild since the matrices
0 I X
0 I Y
0 0 I and 0 0 I
0 0 0
0 0 0
are unitarily similar if and only if X and Y are unitarily similar (see also
[14]).
(ii) Let us consider the problem of classifying m-tuples (p1 , . . . , pm ) of
projectors p2i = pi in a unitary space.
12
For m = 1 this problem is unitarily time; the canonical matrix of a
projector p = p2 was obtained in [15] and [16]. Of course, it is
I D
,
0 0
where D is of the form (6), since a matrix F = [Fij ] of the form (7) satisfies
F 2 = F only if k = 2, F11 = I, and F22 = 0.
As was proved in [16], for m ≥ 2 this problem is unitarily wild even if
p1 is an orthoprojector, i.e. p1 = p21 = p∗1 , (since the pairs of idempotent
matrices
I 0
X I −X
,
0 0
X I −X
and
I 0
0 0
Y
,
Y
I −Y
I −Y
are unitarily similar if and only if X and Y are unitarily similar), or if
p1 p2 = p2 p1 = 0.
(iii) The problems of classifying the following operators and systems of
operators in unitary spaces are unitarily wild:
• Pairs of linear operators (ϕ, ψ) such that
ϕ2 = ψ 2 = ϕψ = ψϕ = 0
since
0 I
0 0
0 X
,
0 0
and
0 I
0 0
0 Y
,
0 0
are unitarily similar if and only if X and Y are unitarily similar.
• Pairs of selfadjoint operators (ϕ, ψ) because ϕ + iψ is an arbitrary
operator. The tame-wild dichotomy for satisfying quadratic relation
pairs of selfadjoint operators in a Hilbert space was studied in [17].
• Pairs of unitary operators (ϕ, ψ) since
(i(ϕ + 1)(ϕ − 1)−1 , i(ψ + 1)(ψ − 1)−1 )
is a pair of selfadjoint operators (the Cayley transformation).
13
• Partial isometries (i.e., linear operators ϕ such that (ϕ∗ ϕ)2 = ϕ∗ ϕ), it
was proved in [18].
(iv) The problem of classifying unitary representations of a connected
quiver Q is unitarily tame if Q ∈ {•, • → •} and unitarily wild otherwise.
Indeed, the classification of unitary representations of the quiver • → • is
given by the singular value decomposition (Lemma 2.1).
The problem of classifying unitary representations of the quiver • → • ← •
is unitarily wild because it reduces to the unitary matrix problem for marked
block matrices of the form
, and two block matrices
♣
2I 0 0 ♣I♣ ♣ ♣ ♣♣♣♣♣ X
♣♣♣♣
0 I 0 ♣I♣ ♣ ♣ ♣♣♣♣ ♣I♣ ♣ ♣
0 0 0 I 0
♣
and
2I 0 0 ♣I♣ ♣ ♣ ♣♣♣♣♣ Y
♣♣♣♣
0 I 0 ♣I♣ ♣ ♣ ♣♣♣♣ ♣I♣ ♣ ♣
0 0 0 I 0
are equivalent if and only if X and Y are unitarily similar. We can change
the direction of an arrow in a quiver by replacing in each representation the
corresponding linear mapping by the adjoint one.
(v) Let us consider the problem of classifying n-tuples (V1 , . . . , Vn ) of
subspaces of a unitary space U up to the following equivalence:
(V1 , . . . , Vn ) ∼ (V1′ , . . . , Vn′ )
if there exists an isometry ϕ : U → U such that
ϕV1 = V1′ , . . . , ϕVn = Vn′ .
Fixing an orthogonal basis in U and (non-orthogonal) bases in V1 , . . . , Vn , we
reduce it to the canonical form problem for block matrices A = [A1 | . . . |An ]
(the columns of Ai are the basis vectors of Vi and hence are linearly independent) up to unitary transformations of rows of A and elementary (nonunitary) transformations of columns of Ai (i = 1, . . . , n). If n = 1, then
A = [A1 ] reduces to I ⊕ 0p0; this follows from Lemma 2.1.
If n = 2, then A = [A1 |A2 ] reduces to
♣
I
0
I ♣♣♣♣♣ ♣0♣ ♣ ♣
♣♣
♣ 0♣ ♣ ♣ ♣♣ D
♣♣♣♣ ,
0 I
0 0
14
where D is of the form (6). This block matrix reduces to a block direct sum
of matrices
1 αi
(α > 0), [1|1], [1|010 ], [010 |1], [010 |010 ].
0 1
(The problem of classifying pairs of subspaces in a complex or real vector
space with scalar product given by a symmetric, or skew-symmetric, or Hermitian form was solved in [19].)
For n = 3 this problem is unitarily wild even if we restrict our consideration to the triples (V1 , V2 , V3 ) with V1 ⊥ V2 since
I 0 X
I 0 X′
0 I Y reduces to 0 I Y ′
0 0 I
0 0 I
if and only if (X, Y ) and (X ′ , Y ′ ) determine isometric unitary representations
of the quiver • ← • → • (see page 14). An analogous statement was proved in
[20] and [13]: the problem of classifying triples (p1 , p2 , p3 ) of orthoprojectors
pi = p2i = p∗i in a unitary space is unitarily wild even if
p1 p2 = p2 p1 = 0;
such a triple determines, in one-to-one manner, a triple (V1 , V2 , V3 ) with
V1 ⊥ V2 by means of Vi = Im pi .
3
Unitary representations of a quiver
From now on, Q denotes a quiver with vertices 1, . . . , p and arrows α1 , . . . , αq .
A unitary representation of dimension
d = (d1 , . . . , dp ) ∈ Np0
(in short, a unitary d representation) will be given by assigning a matrix
Aα ∈ Cdj ×di to each arrow α : i → j, i.e., by the sequence
A = (Aα1 , . . . , Aαq )
(assigning to each vertex i the unitary vector space Cdi with scalar product
(x, y) = x̄1 y1 + · · · + x̄di ydi ,
15
we obtain a unitary representation; see page 2).
∼
An isometry A → B of d representations A and B (an autometry if
A = B) is given by a sequence
S = (S1 , . . . , Sp )
of unitary di × di matrices Si such that
Sj Aα = Bα Si
for each arrow α : i → j; we say also that B is obtained from A by admissible
transformations and write A ≃ B. An autometry S : A→A
˜ is scalar if
S = a1A , where
a ∈ C, 1A = (Id1 , . . . , Idp ).
For two sequences of matrices
M = (M1 , . . . , Mt ),
N = (N1 , . . . , Nt ),
we denote
M ⊕ N = (M1 ⊕ N1 , . . . , Mt ⊕ Nt ).
A unitary d representation A of Q is indecomposable if (i) d 6= (0, . . . , 0) and
(ii) A ≃ B ⊕ C implies B or C has dimension (0, . . . , 0).
3.1
Canonical representations
Let A be a unitary representation of Q. Using the algorithm from page 6, we
reduce Aα1 to its canonical form A∞
α1 , then restrict the set of admissible transformations with A to those that preserve A∞
α1 (it gives certain unitary matrix
problems for Aα2 , . . . , Aαq with partitions them into blocks) and reduce Aα2
to its canonical form A∞
α2 , and so on. The obtained representation
∞
A∞ = (A∞
α1 , . . . , Aαq )
(we omit the marking lines) will be called a canonical representation of the
quiver Q; the sequence of the schemes
∞
S(A∞ ) = (S(A∞
α1 ), . . . , S(Aαq ))
will be called the scheme of A∞ .
Clearly, A ≃ A∞ and A ≃ B if and only if A∞ = B ∞ .
16
Theorem 3.1. (a) Every unitary representation is isometric to a representation of the form
B = (P1 ⊗ Im1 ) ⊕ · · · ⊕ (Pt ⊗ Imt )
≃ P1 ⊕ · · · ⊕ P1 ⊕ · · · ⊕ Pt ⊕ · · · ⊕ Pt
|
|
{z
}
{z
}
m1 copies
mt copies
(see (9)), where P1 , . . . , Pt are nonisometric indecomposable representations,
uniquely determined up to isometry, and m1 , . . . , mt are uniquely determined
∼
natural numbers. Every autometry S : B → B has the form
S = (1P1 ⊗ U1 ) ⊕ · · · ⊕ (1Pt ⊗ Ut ),
where Ui is a unitary mi × mi matrix (1 ≤ i ≤ t).
(b) A unitary representation of dimension 6= (0, . . . , 0) is indecomposable
if and only if all its autometries are scalar.
Proof. Analogously (5), the matrices of every unitary representation A of Q
can be accommodated in a block diagonal matrix
A = diag(Aαq , Aαq−1 , . . . , Aα1 , 0, . . . , 0)
with a certain set of marked blocks such that the admissible transformations with A correspond to the admissible transformations with M(A). Then
M(A∞ ) = M(A)∞ and we can apply Theorem 2.2.
3.2
The set of dimensions of indecomposable unitary
representations
We will use the following notation:
• MQ = [mij ] is the p × p matrix, in which mij is the number of arrows
i → j and i ← j of the quiver Q;
• supp(z) is the full subquiver of Q with the vertex set {i | zi 6= 0} for
each z ∈ Np0 ;
• ei = (0, . . . , 1, . . . , 0) ∈ Np0 with 1 in the ith position.
17
Denote by D(Q) the subset of Np0 consists of e1 , . . . , ep , all ei + ej with
mij = 1, and all nonzero z with connected supp(z) ∈
/ {•, • → •} such that
zMQ ≥ z.
In this section we prove:
Theorem 3.2. D(Q) is the set of dimensions of indecomposable unitary
representations of a quiver Q.
Put
∆i (z) :=
X
z ∈ Np0 ,
mij zj ,
j
1 ≤ i ≤ p.
Then
zMQ = (∆1 (z), . . . , ∆p (z)).
Lemma 3.1. D(Q) satisfies the following conditions:
(i) If z ∈ D(Q) and supp(z) ∈
/ {•, • → •, • ❨✁}, then zMQ > z.
(ii) If z, u ∈ D(Q) and z < u, then there exists i such that z + ei ≤ u and
z + ei ∈ D(Q).
Proof. (i) Let
z ∈ D(Q),
supp(z) ∈
/ {•, • → •, • ❨✁},
zMQ = z.
Fix i such that zi = max{z1 , . . . , zp }. Then mij 6= 0 for a certain j 6= i. Since
zj = ∆j (z) ≥ mij zi ≥ zi ,
we have
zi = zj ,
mij = 1,
mkj zk = 0
for all k 6= i. Taking zj and zi instead of zi and zj , we obtain mki zk = 0 for
all k 6= j. Hence supp(z) = • → •, a contradiction.
(ii) Let z, u ∈ D(Q) and z < u. If supp(z) 6= supp(u), then there exists
a nonzero mij with
i ∈ supp(u) \ supp(z),
j ∈ supp(u) ∩ supp(z).
The z + ei satisfies the requirements.
18
We may assume that
supp(z) = supp(u) = Q.
Then Q 6∈ {•, • → •}. Fix a vertex l such that zl < ul . We will suppose that
∆l (z) = zl and mll = 0 (otherwise z + el satisfies the requirements).
Assume first that zl ≤ zj for some mlj 6= 0. The condition ∆l (z) = zl
implies zl = zj , mlj = 1, and mlk = 0 for all k 6= j. Hence
zj = zl < ul ≤ ∆l (u) = uj .
Since Q 6= • → •, mjk 6= 0 for some k 6= l, and we can take z + ej .
Next, let zl > zj (and hence z + ej ∈ D(Q)) for all nonzero mlj . If zj = uj
for all mlj 6= 0, then
ul ≤ ∆l (u) = ∆l (z) = zl ,
a contradiction. Hence zj < uj for a certain mlj 6= 0, and we can take
z + ej .
Lemma 3.2. If A is a unitary d representation of Q and d ∈
/ D(Q), then A
is decomposable.
Proof. Assume to the contrary, that A is indecomposable. Then supp(d) is
connected; Lemma 2.1 and d ∈
/ D(Q) imply
supp (d) ∈
/ {•, • → •},
dMQ d,
that is, there exists l such that ∆l (d) < dl . Then mll = 0 and we can assume
that there are no arrows starting from l (otherwise we replace each arrow
α : l → i by α∗ : i → l, simultaneously replacing Aα by the adjoint matrix).
Let α, β, . . . , γ be all arrows stopping at l; combine the corresponding
them matrices of A into a single dl × ∆l (d) matrix
[Aα |Aβ | · · · |Aγ ].
The number of its rows is greater than the number of its columns; making a
zero row by unitary transformations of rows, we obtain A ≃ B ⊕ P , where
P is the zero representation of dimension el , a contradiction.
19
Lemma 3.3. If there exists an indecomposable unitary z representation and
zi < ∆i (z) for a certain vertex i, then there exists an indecomposable unitary
z + ei representation.
Proof. Let A be an indecomposable unitary z representation and z1 < ∆1 (z).
We can assume that each starting from the vertex 1 arrow is a loop (replacing
each λ : 1 → j, j 6= 1, by λ∗ : j → 1 and, respectively, Aλ by A∗λ∗ ).
1) Assume first that there is a loop α : 1 → 1 and define a unitary z + e1
representation H in which
• Hα is the nilpotent Jordan block of size (z1 + 1) × (z1 + 1),
• Hβ := Aβ ⊕ 011 for each β : 1 → 1, β 6= α;
• Hγ := Aγ ⊕ 010 for each γ : j → 1, j 6= 1; and
• Hδ := Aδ for each δ : j → k, k 6= 1.
The representation H is indecomposable.
Indeed, let A− and H − denote the restrictions of A and H on the subquiver Q− := Q \ α. By Theorem 3.1(a), we may assume that
A− = (P1 ⊗ Im1 ) ⊕ · · · ⊕ (Pt ⊗ Imt ),
where P1 , . . . , Pt are nonisometric indecomposable representations of Q− , P1
is the zero representation of dimension e1 , and
m1 ≥ 0, m2 > 0, . . . , mt > 0.
Clearly,
H − = (P1 ⊗ Im1 +1 ) ⊕ (P2 ⊗ Im2 ) ⊕ · · · ⊕ (Pt ⊗ Imt ).
Let
∼
∼
S = (S1 , S2 , . . . ) : H → H.
Since S : H − → H − , by Theorem 3.1(a)
S = (1P1 ⊗ U1 ) ⊕ · · · ⊕ (1Pt ⊗ Ut ),
(d11 )
S1 = U1
(d1t )
⊕ · · · ⊕ Ut
20
,
(12)
where
(d1j , . . . , dpj ) = dim(Pj )
and Uj is a unitary matrix (1 ≤ j ≤ t). Since
S1 H α = H α S1 ,
Hα is a Jordan block and S1 is a unitary matrix, we have S1 = aI, a ∈ C. The
representation A is indecomposable, so that d1j 6= 0 and by (12) Uj = aI for
all 1 ≤ j ≤ t. Hence S = a1H and H is indecomposable by Theorem 3.1(b).
2) There remains the case m11 = 0. Let
α1 : j1 → 1, . . . , αl : jl → 1
be all the arrows stopping at 1. We denote by A− the restriction of A on the
subquiver
Q− := Q \ {1; α1 , . . . , αl }.
By Theorem 3.1(a), we may assume that
A− = (P1 ⊗ Im1 ) ⊕ · · · ⊕ (Pt ⊗ Imt ),
where P1 , . . . , Pt are nonisometric indecomposable unitary representations of
Q− .
∼
Let (S2 , . . . , Sp ) : A− → A− . By Theorem 3.1(a),
Si = (Idi1 ⊗ U1 ) ⊕ · · · ⊕ (Idit ⊗ Ut ),
where (d2j , . . . , dpj ) = dimPj . For an arbitrary unitary z1 × z1 matrix S1 , we
define à by means of
∼
S = (S1 , S2 , . . . , Sp ) : A → Ã
(then Ã− = A− ). Taking into account that
Ãατ = S1−1 Aατ Sjτ
and partitioning the sets of columns of every Aατ and Ãατ in the same manner
as Sjτ , we obtain
B := [Aα1 | · · · |Aαl ] = [B1 | · · · |Bm ]
21
and
B̃ := [Ãα1 | · · · |Ãαl ] = [B̃1 | · · · |B̃m ],
where
m=
l
X
τ =1
(djτ 1 + · · · + djτ t )
and
B̃i = S1−1 Bi Uf (i)
for a certain f (i) ∈ {1, . . . , t}.
Let z1 × ui be the size of Bi and put
ri = rank[B1 | · · · |Bi−1 |Bi+1 | · · · |Bm ].
B is a z1 × ∆1 (z) matrix and z1 < ∆1 (z), so z1 − ri < ui for a certain i. Since
S1 and Uf (i) are arbitrary unitary matrices, by Lemma 2.1(a) there exists S
such that
C1 · · · Ci−1 Ci Ci+1 · · · Cm
B̃ =
,
0 ···
0
D
0
···
0
where the rows of
[C1 | · · · |Ci−1 |Ci+1 | · · · |Cm ]
are linearly independent and D is a (z1 − ri ) × ui matrix of the form
diag(a1 , . . . , an ) ⊕ 0kh
with real a1 ≥ · · · ≥ an > 0. Since à is indecomposable and z1 − ri < ui , we
have k = 0 and h > 0.
Let an+1 be a real number such that an > an+1 > 0. The replacement D
by
D ′ = diag(a1 , . . . , an , an+1 ) ⊕ 00,h−1
changes B̃ to a new matrix B̃ ′ and à to a new representation H of dimension
z + e1 .
∼
Let R : H → H. Since H and A coincide on Q− and
∼
(R2 , . . . , Rp ) : A− → A− ,
by Theorem 3.1(a) the matrices R2 , . . . , Rp have the form
Rj = (Idj1 ⊗ V1 ) ⊕ · · · ⊕ (Idjτ ⊗ Vτ )
22
with unitary V1 , . . . , Vτ . By
R1−1 B̃ ′ (Rj1 ⊕ · · · ⊕ Rjl ) = B̃ ′ ,
R1 has the form R11 ⊕ R12 , where
−1 ′
R12
D Vf (i) = D ′ .
Lemma 2.1 implies
R12 = R13 ⊕ [c].
Putting
R̃1 = R11 ⊕ R13 ,
∼
R̃j = Rj (j > 1),
we have R̃ : Ã → Ã. By Theorem 3.1(b),
for some a ∈ C, so
R̃j = aI,
1 ≤ j ≤ p,
Vj = aI,
1 ≤ j ≤ τ.
In particular, Vf (i) = aI and, since
−1 ′
R12
D Vf (i) = D ′ ,
c = a and R1 = aI. Therefore R = a1H and H is indecomposable by Theorem
3.1(b).
Proof of Theorem 3.2. Let U(Q) denote the set of dimensions of indecomposable unitary representations of Q. Lemma 3.2 implies U(Q) ⊂ D(Q).
Let u ∈ D(Q). Then ui 6= 0 for a certain i. Using Lemma 3.1(ii), we
select a sequence
u1 := ei , u2 , . . . , ut := u
in D(Q) such that
u2 − u1 , . . . , ut − ut−1 ∈ {e1 , . . . , ep }.
By Lemma 3.3,
{u1, . . . , ut } ⊂ U(Q),
23
D(Q) ⊂ U(Q).
3.3
The number of parameters in an indecomposable
unitary representation
By the number of real (complex) parameters of a unitary representation A
we mean the number of circles (stars) in the scheme S(A∞ ). Recall that
to circles correspond positive real numbers in A∞ , and to stars correspond
complex numbers; the other entries in A∞ are zeros.
Kac [7, Theorem C] proved that the maximal number of paremeters in
an indecomposable (non-unitary) representation of dimension d over an algebraically closed field is 1 − ϕQ (d), where
ϕQ (x) = x21 + · · · + x2p −
p
X
mij xi xj
i,j=1
is a Z-bilinear form called the Tits form of the quiver Q, and mij is the
number of arrows i → j and i ← j.
We say that a zone (see page 9) is in general position if all its diagonal
entries are distinct and, if it is an equivalence zone, nonzero. A unitary
representation A is said to be in general position if all zones in A∞ are in
general position.
Theorem 3.3. (a) For every d ∈ D(Q) (see page 18) there exists an indecomposable canonical unitary d representation of general position, its scheme
is uniquely determined by d.
P
(b) An indecomposable unitary d representation A has
di − 1 real parameters and at most
1X
1 − ϕQ (d) +
di (di − 1)
2
complex parameters; this number is reached if and only if A is in general
position.
Proof. We consider the set of zones of a canonical unitary representation
∞
A∞ = (A∞
α1 , . . . , Aαq ) as linearly ordered:
Z1 < Z2 if i1 < i2 ; or i1 = i2 and l1 < l2 ;
or i1 = i2 , l1 = l2 and Bl(Z1 ) < Bl(Z2 )
(13)
(see (2) and Section 2.2); where Zk (k = 1, 2) is a zone of depth lk in A∞
αi .
k
24
(a) Let d ∈ D(Q). By Theorem 3.2, there exists an indecomposable unitary d representation A. Let A be not in general position, and let Z be
the first (in the sense of (13)) zone of A∞ that is not in general position.
Changing diagonal entries of Z, we transform it into a zone Z̃ of general
position and A∞ into a new representation Ã. This exchange narrows down
the set of admissible transformations that preserve all zones ≤ Z, and, by
Theorem 3.1(b), A∞ has only scalar autometries (as an indecomposable representation), therefore, Ã has only scalar autometries and is indecomposable
too.
If à is not in general position, we repeat this process for it, and so on,
until we obtain an indecomposable d representation B of general position.
Its scheme is uniquely determined since, for each zone Z of B, the set of
admissible transformations that preserve all zones ≤ Z (and hence the matrix
problem for the remaining part of B) does not depend on diagonal entries of
Z such that it is in general position.
(b) Let A be an indecomposable canonical d representation, and Z be its
zone or the symbol ∞. Denote by J(Z) the set of all isometries of the form
∼
S : A → Ã that preserve all zones < Z (all zones if Z = ∞). As follows
from the algorithms from pages 6 and 16, J(Z) consists of all sequences of
the form S = (S1 , . . . , Sp ), where
Si = Uσ(i1) ⊕ Uσ(i2) ⊕ · · · ⊕ Uσ(iti ) ,
σ : {(ij) | 1 ≤ i ≤ p, 1 ≤ j ≤ ti } → {1, . . . , t}
is a fixed surjection, and U1 , . . . , Ut are arbitrary unitary matrices of fixed
sizes m1 × m1 , . . . , mt × mt (we will write S = S(U1 , . . . , Ut )).
Put
∆1 (Z) = m1 + · · · + mt , ∆2 (Z) = m21 + · · · + m2t .
Let Z 6= ∞ and Z ′ be the first zone after Z (Z ′ = ∞ if Z is the last zone
of A). We will prove that
∆1 (Z) − ∆1 (Z ′ ) = n• (Z),
∆2 (Z) − ∆2 (Z ′ ) ≤ 2n(Z) − n• (Z) − 2n⋆ (Z),
(14)
(15)
and that the equality in (15) holds if and only if Z is a zone of general
position; where n(Z) is the number of entries in Z, and n• (Z) (resp., n⋆ (Z))
is the number of circles (resp., stars) that correspond to the diagonal entries
of Z.
25
As follows from the algorithms from pages 6 and 16, the block Bl(Z) is
reduced by transformations
Bl(Z) 7→ Ui−1 Bl(Z)Uj ,
(16)
where
S = S(U1 , . . . , Ut ) ∈ J(Z)
and i and j are determined by Z; moreover, this S is contained in J(Z ′ ) if
and only if (16) preserves Z.
(i) Let i 6= j, say, i = 1 and j = 2. Then, by Lemma 2.1,
Z = Bl(Z) = a1 Ir1 ⊕ · · · ⊕ ak−1 Irk−1 ⊕ 0xy , rα ≥ 1,
x ≥ 0, and y ≥ 0. The transformation (16) preserves Z if and only if
U1 = V1 ⊕ · · · ⊕ Vk and U2 = V1 ⊕ · · · ⊕ Vk−1 ⊕ Vk+1 ,
(17)
where V1 , . . . , Vk+1 are unitary matrices of sizes
r1 × r1 , . . . , rk−1 × rk−1 , x × x, y × y.
Hence, J(Z ′ ) consists of all S ∈ J(Z) with U1 and U2 of the form (17), that
is,
S = S(V1 , . . . Vk+1, U3 . . . Ut ).
Therefore,
∆1 (Z ′ ) = r1 + · · · + rk−1 + x + y + m3 + · · · + mt ,
2
∆2 (Z ′ ) = r12 + · · · + rk−1
+ x2 + y 2 + m23 + · · · + m2t .
By (16), Bl(Z) has size m1 × m2 ,
m1 = r1 + · · · + rk−1 + x,
m2 = r1 + · · · + rk−1 + y,
so
n(Z) = m1 m2 ,
n• (Z) = r1 + · · · + rk−1 ,
n⋆ (Z) = 0.
We have
∆1 (Z) − ∆1 (Z ′ ) = r1 + · · · + rk−1 = n• (Z)
26
and
∆2 (Z) − ∆2 (Z ′ ) = (r1 + · · · + rk−1 + x)2
2
+ (r1 + · · · + rk−1 + y)2 − r12 − · · · − rk−1
− x2 − y 2
= [(r1 + · · · + rk−1 + x) − (r1 + · · · + rk−1 + y)]2
+ 2(r1 + · · · + rk−1 + x)(r1 + · · · + rk−1 + y)
2
− r12 − · · · − rk−1
− x2 − y 2
2
= (x − y)2 + 2n(Z) − r12 − · · · − rk−1
− x2 − y 2
2
= −2xy + 2n(Z) − r12 − · · · − rk−1
≤ 2n(Z) − r1 − · · · − rk−1 = 2n(Z) − n• (Z).
Moreover, we have the equality if and only if
r1 = · · · = rk−1 = 1,
xy = 0,
i.e, Z is in general position.
(ii) Let i = j, say, i = j = 1. Then, by Lemma 2.2, Bl(Z) = [Fαβ ], where
Fαβ = 0 if α > β, and
Fαα = λα Irα , rα ≥ 1, r1 + · · · + rk = m1 .
The transformation (16) preserves
Z = {Fαβ | α ≤ β}
if and only if
U1 = V1 ⊕ · · · ⊕ Vk ,
where V1 , . . . , Vk are unitary matrices of sizes
r 1 × r 1 , . . . , rk × r k .
Hence, J(Z ′ ) consists of all S ∈ J(Z) with
U1 = V1 ⊕ · · · ⊕ Vk ,
that is,
S = S(V1 , . . . , Vk , U2 , . . . , Ut ).
So
∆1 (Z) − ∆1 (Z ′ ) = m1 − r1 − · · · − rk = 0 = n• (Z)
27
and
∆2 (Z) − ∆2 (Z ′ ) = (r1 + · · · + rk )2 − r12 − · · · − rk2
X
rα rβ − 2(r12 + · · · + rk2 )
=2
α≤β
= 2n(Z) − 2(r12 + · · · + rk2 )
≤ 2n(Z) − 2(r1 + · · · + rk )
= 2n(Z) − 2n⋆ (Z).
Moreover, we have the equality if and only if
r1 = · · · = rk = 1,
that is, Z is in general position.
Hence, the relations (14) and (15) hold.
Let Z1 < · · · < Zr be all zones of A ordered by (13), and let
dim A = (d1 , . . . , dp ).
Then Zi′ = Zi+1 for i < r, and Zr′ = ∞. Since J(Z1 ) consists of all sequences
S = (S1 , . . . , Sp ) of unitary d1 × d1 , . . . , dp × dp matrices,
∆2 (Z1 ) = d21 + · · · + d2p .
∆1 (Z1 ) = d1 + · · · + dp ,
Since A is indecomposable, by Theorem 3.1(b) J(∞) consists of all sequences
λ ∈ C,
S = λ(Id1 , . . . , Idp ),
|λ| = 1,
so
S = S([λ]),
∆1 (∞) = ∆2 (∞) = 1.
By (14),
d1 + · · · + dp − 1 = ∆1 (Z1 ) − ∆1 (∞)
r
r
X
X
′
=
(∆1 (Zi ) − ∆1 (Zi )) =
n• (Zi )
i=1
i=1
is the number of circles in S(A∞ ), that is, the number of real parameters in
A.
28
By (15),
d21 + . . . + d2p − 1 = ∆2 (Z1 ) − ∆2 (∞)
r
X
(∆2 (Zi ) − ∆2 (Zi′ ))
=
i=1
r
X
≤2
But
Pr
i=1
i=1
n(Zi ) −
r
X
i=1
n• (Zi ) − 2
r
X
n⋆ (Zi ).
i=1
n(Zi ) is the number of entries in Aα1 , . . . , Aαq , hence, it is equal to
p
X
mij di dj ,
i,j=1
where mij is the number of arrows i → j and i ← j;
⋆
n (A) :=
r
X
n⋆ (Zi )
i=1
is the number of complex parameters in A. Therefore,
X
1 X 2
1 X
n⋆ (A) ≤
mij di dj − (
di − 1) − (
di − 1)
2
2
X
X
1X 2
(di − di)
= 1−[
d2i −
mij di dj ] +
2
1X
= 1 − ϕQ (d) +
di (di − 1).
2
We have the equality if and only if all Zi are in general position, i.e., A is in
general position.
The proof implies
Corollary 3.1. (a) Let d ∈ D(Q) and m = max{d1 , . . . , dp }. Then there
exists an indecomposable canonical d representation of general position with
entries in {0, 1, . . . , m}.
P
(b) A decomposable unitary d representation has less than
di − 1 real
parameters and less than
1X
di (di − 1)
1 − ϕQ (d) +
2
complex parameters.
29
Proof. (a) This statement follows from Theorems 3.3(a) and 2.3.
(b) This statement is proved as Theorem 3.3(b), but, in the last two
paragraphs of its proof, we must use ∆1 (∞) > 1 and ∆2 (∞) > 1 instead of
∆1 (∞) = ∆2 (∞) = 1 since J(∞) contains a non-scalar authometry in the
case of a decomposable representation A.
4
Euclidean representations of a quiver
Let Q be a quiver with vertices 1, . . . , p and arrows α1 , . . . , αq . A Euclidean
representation A of dimension
d = (d1 , . . . , dp ) ∈ Np0
will be given by assigning a matrix Aα ∈ Rdj ×di to each arrow α : i → j; i.e.,
by the sequence
A = (Aα1 , . . . , Aαq ).
∼
An R-isometry A →R B of Euclidean representations A and B will be given
by a sequence S = (S1 , . . . , Sp ) of real orthogonal matrices such that Sj Aα =
∼
Bα Si for each arrow α : i → j (analogously, R : A →C B denotes an isometry
in the sense of Section 3). A Euclidean d representation A is said to be Rindecomposable if (i) d 6= (0, . . . , 0) and (ii) A ≃R B ⊕ C implies that B or C
has dimension (0, . . . , 0).
For a sequence of complex matrices
M = (M1 , . . . , Mn ),
we define the conjugate sequence
M̄ = (M̄1 , . . . , M̄n ),
the transposed sequence
M T = (M1T , . . . , MnT ),
and the adjoint sequence
M ∗ = M̄ T .
Clearly, the Euclidean representations are the selfconjugate unitary representations.
30
4.1
A reduction to unitary representations
We give a standard reduction of the problem of classifying Euclidean representations to the problem of classifying unitary representations.
Let ind(Q) and indR (Q) denote complete systems of nonisometric indecomposable unitary representations and non-R-isometric R-indecomposable
Euclidean representations respectively. Let us replace each representation in
ind(Q) that is isometric to a Euclidean representation by a Euclidean one,
∼
and denote the set of such by ind0 (Q) (if A ∈ ind(Q) and S : A→C Ā, then A
is isometric to a Euclidean representation if and only if S T = S; see Theorem
4.2). Denote by ind1 (Q) the set consisting of all representations from ind(Q)
that are isometric to their conjugates, but not to a selfconjugate, together
with one representation from each pair {A, B} ⊂ ind(Q) such that
A6≃C Ā ≃C B.
For a unitary d representation
A = (Aα1 , . . . , Aαq ),
we define the Euclidean 2d representation
AR = (ARα1 , . . . , ARαq ),
where ARα , is obtained from Aα by replacing each entry a + bi (a, b ∈ R) by
the block
a b
−b a
Since
U
with the unitary
−1
a + bi 0
a b
U=
0 a − bi
−b a
1 1 −1
U=√
,
2 i i
we have
AR ≃C A ⊕ Ā.
31
(18)
Theorem 4.1. (a) Let A and B be Euclidean representations of a quiver Q.
Then A ≃R B if and only if A ≃C B.
(b) Every Euclidean representation is R-isometric to a direct sum of indecomposable Euclidean representations, uniquely determined up to R-isometry
of summands. Moreover,
indR (Q) = ind0 (Q) ∪ {AR | A ∈ ind1 (Q)}.
(19)
(c) The set of dimensions of R-indecomposable Euclidean representations
of Q coincides with the set of dimensions of indecomposable unitary representations and is equal to D(Q) (see page 18).
A homomorphism (R-homomorphism) S : A → B of representations A
and B of Q is a sequence of complex (real) matrices
S = (S1 , . . . , Sp )
such that Sj Aα = Bα Si for each arrow α : i → j. Clearly, an isomorphism S
is an isometry if and only if S ∗ = S −1 .
Lemma 4.1. The following properties are equivalent for a unitary
(Euclidean) representation A:
(i) A is decomposable (R-decomposable).
(ii) There exists an endomorphism (R-endomorphism) F : A → A such
that F = F ∗ = F 2 ∈
/ {0A , 1A }.
(iii) There exists a nonscalar selfadjoint endomorphism (Rendomorphism) S = S ∗ : A → A.
∼
Proof. (i)⇒(ii) Let S : A → B ⊕ C be an isometry of unitary representations
(R-isometry of Euclidean representations) and B 6= 0 6= C. Then
F := S −1 (1B ⊕ 0C )S : A → A
satisfies (ii).
(ii)⇒(iii) Put S := F .
(iii)⇒(i) Since every Si in S is a Hermitian (resp., real symmetric) matrix,
there exists a unitary (real orthogonal) matrix Ui such that
Ri := Ui Si Ui−1 = diag(ai1 , . . . , aiti ),
32
where aij ∈ R and
ai1 ≥ · · · ≥ aiti .
Define the unitary (Euclidean) representation B by means of the isometry
∼
U := (U1 , . . . , Up ) : A → B.
Then
∼
R := USU −1 : B → B
is an isometry and
R = a1 I1 ⊕ · · · ⊕ at It ,
where
a1 > · · · > at ,
Ii = (Ini1 , . . . , Inip ),
nij ≥ 0.
Clearly,
B = B1 ⊕ · · · ⊕ Bt ,
where
dim(Bi ) = (ni1 , . . . , nip ).
Proof of Theorem 4.1. 1) We first prove the statement (a) for an Rindecomposable Euclidean representation A. Let
∼
S = Φ + iΨ : A →C B,
where Φ and Ψ are real matrices and B is a Euclidean representation. Then
Φ and Ψ are R-homomorphisms A → B. Since
1A = S ∗ S = (ΦT − iΨT )(Φ + iΨ)
= (ΦT Φ + ΨT Ψ) + i(ΦT Ψ − ΨT Φ),
we have
ΦT Φ + ΨT Ψ = 1A .
By Lemma 4.1, the selfadjoint R-endomorphisms ΦT Φ and ΨT Ψ are scalar,
that is,
ΦT Φ = λ1A , ΨT Ψ = µ1A , λ + µ = 1.
33
Obviously, λ and µ are non-negative real numbers. For definiteness, λ > 0,
1
∼
then λ− 2 Φ : A →R B.
2) Let A be an R-indecomposable Euclidean representation that is decomposable as a unitary representation. We prove that
A ≃R B R ≃C B ⊕ B̄,
where B is an indecomposable unitary representation that is not isometric
to a Euclidean representation.
Indeed, by Lemma 4.1 there exists an endomorphism F : A → A such
that
F = F∗ = F2 ∈
/ {0A , 1A }.
Let F = Φ + iΨ, where Φ and Ψ are sequences of real matrices. Since
F = F ∗ = ΦT − iΨT ,
it follows that Φ = ΦT and Ψ = −ΨT . By Lemma 4.1, the endomorphism Φ
is scalar, i.e., Φ = λ1A , λ ∈ R. If λ = 0, then
iΨ = F = F 2 = −Ψ2
and Ψ = 0A , a contradiction.
Hence λ 6= 0. Since
F = F 2 = (λ1A + iΨ)2 = (λ2 1A − Ψ2 ) + 2λiΨ,
we have
λ2 1A − Ψ2 = λ1A ,
2λΨ = Ψ.
1
λ= ,
2
1
Ψ2 = − 1A .
4
The condition F 6= 1A implies
Ψ 6= 0A ,
By [21, Sect. 4.4, Exercise 25], every nonsingular skew-symmetric real
matrix is real orthogonally similar to a direct sum of matrices of the form
0 a
,
a > 0.
−a 0
Since
1
Ψ2 = − 1A ,
4
ΨT = −Ψ,
34
there exists a sequence S of real orthogonal matrices such that
1
0 1
−1
SΨS = I ⊗
−1 0
2
(see (9)), where I = (I, . . . , I). Put
G := SF S
−1
1 i
1
.
= I⊗
2
−i 1
∼
Define the Euclidean representation C by means of S : A →R C. Then
G : C → C is an R-endomorphism. It follows from the form of G and the
definition of homomorphisms, that C = B R for a certain B.
If B is a decomposable unitary representation, say, B ≃C X ⊕ Y , then by
(18)
A ≃R B R ≃C B ⊕ B̄ ≃C X ⊕ Y ⊕ X̄ ⊕ Ȳ ≃C X R ⊕ Y R ,
by 1) A ≃R X R ⊕ Y R , a contradiction.
If B is isometric to a Euclidean representation, say,
B ≃C D = D̄,
then
A ≃R B R ≃C B ⊕ B̄ ≃C D ⊕ D,
by 1) A ≃R D ⊕ D, a contradiction. This proves 2).
(a)–(b). Let A and B be Euclidean representations, A ≃R B,
A ≃R A1 ⊕ · · · ⊕ Al ,
B ≃R B1 ⊕ · · · ⊕ Br ,
where Ai and Bj are R-indecomposable. From 2) and Theorem 3.1(a), l = r
and, after a permutation of summands, Ai ≃C Bi . By 1), Ai ≃R Bi . The
equality (19) is obvious.
(c). By Corollary 3.1(a), there exists an R-indecomposable Euclidean
representation (with entries in N0 ) of dimension z for every z ∈ D(Q).
Conversely, let A be an R-indecomposable Euclidean representation. If
A is indecomposable as a unitary representation, then by Theorem 3.2
dim(A) ∈ D(Q). Otherwise by 2) A ≃C B ⊕ B̄, where B is an indecomposable unitary representation, i.e., d := dim(B) ∈ D(Q). Since B is not
isometric to a Euclidean representation,
supp (d) ∈
/ {•, • → •}.
35
Applying twice the definition of D(Q) (see page 3.2), we have
dMQ ≥ d,
4.2
2dMQ ≥ 2d,
dim(A) = 2d ∈ D(Q).
Unitary representations that are isometric to Euclidean representations
Theorem 4.1(b) reduces the problem of classifying Euclidean representations
of a quiver Q to the following two problems:
• classify unitary representations of Q (i.e., construct the set ind(Q));
• bring to light for each A ∈ ind(Q) whether it is isometric to a Euclidean
representation and to construct that representation.
In this section we consider the second problem.
Lemma 4.2. (a) If S is a symmetric unitary matrix, then there exists a
unitary matrix U such that S = U T U.
(b) If S is a skew-symmetric unitary matrix, then there exists a unitary
matrix U such that
0 1
0 1
T
U.
⊕···⊕
S=U
−1 0
−1 0
Proof. Analogous statement for a non-unitary matrix S is given in [21, Sect.
4.4, Corollary 4.4.4 and Exercise 26]. The condition of unitarity makes its
proof much more easy. We give it sketchily since an explicit form of U is
needed for the applications of the next theorem.
Given a symmetric (skew-symmetric) unitary matrix Sn with rows
s1 , . . . , sn . If
s1 6= e1 := (1, 0, . . . , 0),
we take a unitary matrix Un with rows u1 , . . . , un such that
u1 = α(e1 + s1 ),
α ∈ C,
(resp., Cu1 + Cu2 = Ce1 + Cs1 ). Then
ū1 Sn = α(e1 + s̄1 )Sn = α(e1 Sn + s̄1 Sn )
= α(s1 + e1 ) = u1 = e1 Un ,
36
hence
(Un−1 )T Sn Un−1 = Ūn Sn Un−1 = [1] ⊕ Sn−1
(resp., then
Ūn Sn Un−1
0 β
⊕ Sn−2 ,
=
−β 0
|β| = 1; replacing u2 by βu2, we make β = 1). If s1 = e1 , we have
Sn = [1] ⊕ Sn−1 ,
Un := In .
We repeat this procedure until we obtain the required
U := Un (I1 ⊕ Un−1 )(I2 ⊕ Un−2 ) · · · (In−1 ⊕ U1 )
(resp., U := Un (I2 ⊕ Un−2 ) · · · ).
Theorem 4.2. (a) Let A be a unitary representation and A 6≃C Ā. Then
A is not isometric to a Euclidean representation.
∼
(b) Let A be an indecomposable unitary representation and S : A →C Ā.
(i) If S = S T , then A is isometric to a Euclidean representation B
∼
given by U : A →C B, where U1 , . . . , Up are arbitrary unitary
matrices such that UiT Ui = Si (they exist by Lemma 4.2(a)).
(ii) If S 6= S T , then S = −S T and A is not isometric to a Euclidean
representation but is isometric to a unitary representation C of
the form
X Y
−Ȳ X̄
∼
given by V : A →C C, where V1 , . . . , Vp are arbitrary unitary
matrices such that
0 I
T
V i = Si
Vi
−I 0
(they exist by Lemma 4.2(b)).
37
∼
Proof. (a) Let R : A →C B, where B is a Euclidean representation. Then
∼
∼
RT = R̄−1 : B̄ →C Ā,
H := RT R : A →C Ā
(observe that H = H T ).
∼
(b) Let A be an indecomposable unitary representation and S : A →C Ā.
∼
Then S̄S : A →C A, by Theorem 3.1(b)
S̄S = λ1A , S = λS̄ −1 = λS T = λ(λS T )T = λ2 S,
and λ ∈ {1, −1}.
∼
(i) Let λ = 1, U : A →C B and U T U = S. Then
∼
U = (U T )−1 S = Ū S : A →C B̄
and B = B̄.
(ii) Let λ = −1. Then A is not isometric to a Euclidean representation
(otherwise, by (a) there exists
∼
H = H T : A →C Ā;
by Theorem 3.1(b)
H T = µS T = −µS = −H,
S −1 H = µ1A ,
∼
a contradiction). Let V : A →C C, where
0 I
T
V i = Si .
Vi
−I 0
Then
∼
V̄ SV −1 : C →C C̄.
If α is an arrow of Q, then
and Cα is of the form
0 I
0 I
Cα = C̄α
−I 0
−I 0
X Y
.
−Ȳ X̄
38
✄
Applying this theorem to unitary representations of the quiver ✯
✂ ♣ , we
obtain
Corollary 4.1. Let A be a complex matrix that is not unitarily similar to a
direct sum of matrices, and let S −1 AS = Ā for a unitary matrix S (such S
exists if A is unitarily similar to a real matrix). Then A is unitarily similar
to a real matrix if and only if S is symmetric.
✷
References
[1] P. Gabriel, Unzerlegbare darstellungen I, Manuscripta Math. 6 (1972)
71–103.
[2] P. Gabriel and A. V. Roiter, Representations of finite-dimensional algebras, Encyclopaedia of Math. Sci., Vol 73 (Algebra VIII), SpringerVerlag, 1992.
[3] D. E. Littlewood, On unitary equivalence, J. London Math. Soc. 28
(1953) 314–322.
[4] V. V. Sergeichuk, Classification of linear operators in a finite dimensional
unitary space, Funkcional. Anal. i Priloz̆en. 18 (no. 3) (1984) 57–62 =
Functional Anal. Appl. 18 (no. 3) (1984) 224–230.
[5] H. Shapiro, A survey of canonical forms and invariants for unitary similarity, Linear Algebra Appl. 147 (1991) 101–167.
[6] V. G. Kac, Infinite root systems, representations of graphs and invariant
theory, Inventiones Math. 56 (1980) 57–92.
[7] V. G. Kac, Infinite root systems, representations of graphs and invariant
theory, II, J. Algebra 78 (1982) 141–162.
[8] V. Kac, Root systems, representations of graphs and invariant theory,
Lecture Notes in Math., vol. 996, Springer-Verlag, 1983, pp. 74–108.
[9] V. V. Sergeichuk, Representations of directed schemes (in Russian), in
Linear Algebra and the Theory of Representations (Ju. A. Mitropol’skiı̆,
Ed.), Inst. Math. Akad. Nauk Ukrain. SSR, Kiev, 1983, pp. 110–134.
39
[10] S. Friedland, Simultaneous similarity of matrices, Adv. Math. 50 (1983)
189–265.
[11] E. N. Klimenko, Classification of linear operators in a 5-dimensional
unitary space, Graduation Thesis, Kiev Univ., 1993.
[12] P. Donovan and M. R. Freislich, The representation theory of finite
graphs and associated algebras, Carleton Math. Lecture Notes 5, Ottawa, 1973.
[13] S. A. Kruglyak and Yu. S. Samoı̆lenko, Structure theorems for families
of idempotents, Ukrainskii Matematicheskii Zhurnal, 50 (no. 4) (1988)
523-533 (in Russian).
[14] Yu. N. Bespalov and Yu. S. Samoı̆lenko, Algebraic operators and pairs of
selfajoint operators connected with a polynomial relation, Funkcional.
Anal. i Priloz̆en. 25 (no. 4) (1991) 72–74 = Functional Anal. Appl. 25
(no. 4) (1991) 289–291.
[15] D. Z̆. Doković, Unitary similarity of projectors, Aequationes Math. 42
(1991) 220–224.
[16] Kh. D. Ikramov, On the canonical form of projectors with respect to
unitary similarity, Zh. Vychisl. Mat. i Mat. Fiz. 36 (no. 3) (1996) 3–5
=Comput. Math. Math. Phys. 36 (no. 3) (1996).
[17] V. L. Ostrovskiı̆ and Yu. S. Samoı̆lenko, Structure theorems for a pair of
unbounded selfajoint operators satisfying quadratic relation, Adv. Sov.
Math. 9 (1992) 131–149.
[18] P. R. Halmos and J. E. McLaughlin, Partial isometries, Pacific J. Math.
13 (1963) 585–596.
[19] V. V. Sergeichuk, Classification of pairs of subspaces in spaces with
scalar product, Ukrain. Mat. Zh. 42 (no. 4) (1990) 549–554 = Ukrainian
Math. J. 42 (no. 4) (1990) 487–491.
[20] S. A. Kruglyak and Yu. S. Samoı̆lenko, Unitary equivalence of sets of
selfajoint operators, Funkcional. Anal. i Priloz̆en. 14 (no. 1) (1980) 60–
62 = Functional Anal. Appl. 14 (no. 1) (1980) 54–55.
40
[21] R. A. Horn and C. R. Johnson, Matrix Analysis, Cambridge U. P., Cambridge, 1985.
41