DOI 10.

1007/s10958-015-2530-2
Journal of Mathematical Sciences, Vol. 209, No. 6, September, 2015
ON BIRATIONAL DARBOUX COORDINATES ON
COADJOINT ORBITS OF CLASSICAL COMPLEX LIE
GROUPS
M. V. Babich∗ UDC 512.643.8, 514.164.1, 517.912

Any coadjoint orbit of the general linear group can be canonically parameterized using an iteration
method, where at each step we turn from the matrix of a transformation A to the matrix of the
transformation that is the projection of A parallel to an eigenspace of this transformation to a
coordinate subspace.
We present a modification of the method applicable to the groups SO(N, C) and Sp(N, C). One
step of the iteration consists of two actions, namely, the projection parallel to a subspace of an
eigenspace and the simultaneous restriction to a subspace containing a co-eigenspace.
The iteration gives a set ofcouples of functions pk , qk on the orbit such that the symplectic
form of the orbit is equal to dpk ∧ dqk . No restrictions on the Jordan form of the matrices
k
forming the orbit are imposed.
A coordinate set of functions is selected in the important case of the absence of nontrivial Jordan
blocks corresponding to the zero eigenvalue, which is the case dim ker A = dim ker A2 . It contains
the case of general position, the general diagonalizable case, and many others. Bibliography: 10
titles.

1. Introduction. Notation. Coordinates on orbits of groups of series A

In this section, we recall the method of canonical parametrization of (co)adjoint orbits of the
general linear group. The method was introduced in [2–4] and suggested by I. M. Gelfand and
M. I. Najmark [1]. In the present paper, the method is extended to matrix groups preserving
a bilinear quadratic form. The existence of such an extension is also based on [1], where
the triangular decompositions of SO(N ) and Sp(N ) were applied to representation theory.
A distinguishing feature of the method is its insensitivity to the Jordan form of matrices
generating the orbit, it is applicable to arbitrary complex orbits.

It must be noted that the assumption A ∈ so(N ) or A ∈ sp(N ) imposes some restrictions
on the Jordan form of A. For example, the rank of any skew-symmetric
⎛ ⎞ matrix is even, hence
0 1 0
none of these Lie algebras contains matrices similar to, say, ⎝0 0 0⎠.
0 0 0
We assume that the Jordan forms of matrices are compatible with the restrictions prescribed
by the Lie algebra under consideration. It is given that matrices belong to the corresponding
Lie algebra.

Another difficulty is more serious. In some cases, the constructed functions are not inde-
pendent, so that they do not form a coordinate set of functions. We need to pick out the
independent functions and combine them into a canonical set. This difficulty can arise only
in the case if the zero root space has nontrivial Jordan blocks.

∗
St.Petersburg Department of Steklov Mathematical Institute, St.Petersburg State University, St.Petersburg,
Russia, e-mail: misha.babich@gmail.com.

Published in Zapiski Nauchnykh Seminarov POMI, Vol. 432, 2015, pp. 36–57. Original article submitted
December 22, 2014.
830 1072-3374/15/2096-0830 ©2015 Springer Science+Business Media New York
 The study of the general case is beyond the scope of the present work. The classification
and birational canonical parametrization of nilpotent orbits is the subject of a separate paper.
Here we will not construct coordinates on such orbits.

The last remark is that the presented formulas are valid in the case of zero eigenvalue. They
do give a birational canonical parametrization of the orbits with complicated zero root space,
but some orbits will not be parameterized. These “missed orbits” are some algebraically closed
subspaces of the orbits already parameterized. A canonical parametrization of such subspaces
is the subject of the theory of Hamiltonian systems with constraints.
There are no significant difficulties in calculating the Jordan form of a parameterized orbit,
this can be done in the same way as for the case of gl(N ). Namely, it suffices to determine the
maximal ranks of all powers of the constructed matrix Pfin over all values of the parameters in
the explicit formula. It is not difficult because Pfin is triangular. The Jordan normal form of
a matrix from the orbit coincides with the form of a matrix J in the Jordan normal form for
which the ranks of all powers are the same as the maximal ranks of the powers of Pfin . We will
not concentrate on this fact either. Before proceeding to our subject, we need to introduce
basic notions and notation.

Let V be an N -dimensional complex linear space. The general linear group GL(N ) acts
on V by changes of bases: (e) → (eF ), where (e) = (e1 , e2 , . . . , eN ), (eF ) = ((eF )1 , (eF )2 , . . . ,
(eF )N ),

ei , (eF )i ∈ V, (eF )k = ei Fik , Fij ∈ C, F ∈ GL(N, C).
i

The algebra gl(N, C) is the space of all matrices. The nondegenerate pairing A, B → tr AB
identifies the algebra with its dual, so we do not distinguish between the algebra gl(N ) and
its dual gl∗ (N ), the adjoint and coadjoint actions of the group. Coadjoint orbits of Lie groups
have been the classical subject of a huge amount of investigations for more than hundred years,
see [5–10]. The manifold of all matrices similar to a given one is isomorphic to its coadjoint
orbit, they are canonically isomorphic symplectic spaces.
Let us denote the (co)adjoint orbit of some element J ∈ gl(N ) by O(J) := ∪F ∈GL(N ) F −1 JF .
The symplectic form on the orbit is denoted by ωJ : TO × TO → C. To introduce the method
of parametrization of the orbit presented in [3, 4], we need more notation.
Notation 1. Let V be represented as a direct sum of two nonzero subspaces: V = L ⊕ M .
The projection of V along L to M is denoted by
||L
ΠM ∈ Hom(V, M ).
Notation 2. Let V be represented as a direct sum of two nonzero subspaces in two special
ways: L1 ⊕ M = L2 ⊕ M = V . The linear transformation of V that sends points of L1 to
points of L2 parallel to M and leaves points of M unchanged is denoted by
||L ||M ||L ||M
ΠM 1 + ΠL2 = (ΠM 2 + ΠL1 )−1 ∈ End V.

We denote finite ordered sets of vectors by boldface letters. For example, a basis (e) can be
split into two parts a and b:
(e) = (e1 , e2 , . . . , eN ) = (a, b), a = e1 , e2 , . . . , en , b = en+1 , en+2 , . . . , eN .
Linear spans of sets of vectors will be denoted by L (. . .); for example, L (a) is the corre-
sponding coordinate subspace.
831
 If the first set a of vectors of a basis belongs to L1 and the other part b of the basis belongs
||L ||M
to M , then the transformations (ΠM 1 + ΠL2 )±1 have block triangular matrices:

||L1 ||M ±1 I 0
(a, b)(ΠM + ΠL2 ) = (a, b) .
∓Q I

The matrix elements of Q form affine coordinates on an algebraically open subset of the
Grassmannian of n-dimensional subspaces of V .

Let λ be an eigenvalue of the matrices from the orbit (i.e., an eigenvalue of J). Consider
ker(A−λ I)∩im(A−λ I)m−1 := L for any m such that L = 0. The space L is well defined by A
and m ∈ N; it is a nonzero subspace of V . We denote the dimension of L by n = n(λ , m) ∈ N;
it is the number of Jordan blocks corresponding to λ of sizes m × m and larger. According to
this definition, a simple eigenvalue corresponds to m = 1.
Let a basis (e) of V be split into two parts (a, b). The subset a consists of n = n(λ , m)
vectors, and the subset b consists of N − n vectors. Consider the new basis (aL , b) :=
||L (a) ||L (b)
(a, b)(ΠL (b) + ΠL ). The coordinate subspace of the new basis is a well-defined subspace L
of an eigenspace of A; consequently, in terms of the basis (aL , b), the transformation A ∈ gl(N )
has a block triangular form, and the original matrix A in terms of the initial basis (a, b) has
the following form:
    −1
I 0 λI P I 0
A= .
Q I 0 A Q I
This representation takes place if A belongs to some algebraically open subset of the orbit.
The functions P : O(J) → Cn×(N −n) , Q : O(J) → C(N −n)×n , A : O(J) → C(N −n)×(N −n) have
been constructed. The following standpoint was presented in [4].

The orbit O(J) is foliated over the Grassmannian. The class of birational trivializations of
this foliation was presented. The base, which is the Grassmannian, was covered by standard
affine maps isomorphic to a linear space, say HG = HG (λ , m), dim HG = n > 0. A fiber of the
foliation has a natural structure of the direct product HG ∗ × O(J), where O(J) is an orbit of

smaller dimension J = J (J, λ , m) and HG ∗ is the linear space dual to H . The direct product
G
∗
HG × HG is a linear space with a natural symplectic structure; consequently, we get a covering
of the orbit O(J) by domains. Each domain is the symplectic space O(J ) × (HG × HG ∗ ), and

the symplectic structures of the initial orbit and the covering naturally agree with each other.
In other words, a linear symplectic space1 has been split off the initial orbit on an algebraically
open set.
Thus the map A → A is a projection of an open subset of O(J) onto an orbit O(J), where
the Jordan form of J is determined by J, λ , and m: the Jordan blocks of J with sizes m and
larger decrease by one, the other Jordan blocks of J and J coincide. The main results of [3, 4]
are as follows:
• The map O(J) → O(J), defined over an algebraically open subset of the orbit, is a
projection of the trivial symplectic fibration.
Its fiber is a linear symplectic space, we denote it by (E, ωE ).
• The couples Pij , Qji of symmetric matrix elements of P, Q are birational Darboux co-
ordinates on the fiber E.

1Hereafter we denote the space H × H ∗ by E .

G G

832
 • The symplectic form ωJ on the orbit is the sum of the forms on the base O(J) and on
the fiber E:
ωJ = ωJ + ωE = ωJ + tr dP ∧ dQ.
For brevity, we skip the notion of pull-backs of projections. The process can be iterated to
exhaust the dimensions of the target orbits. This gives Darboux coordinates on the initial
orbit O(J). If we start with the orbit of a matrix J with zero trace, we get a parametrization
of an orbit of SL(N ) embedded into sl(N ). It is a parametrization of the orbits of the groups
of series A.
Let us consider the dual pattern, which gives us the representation
   −1
I 0 A P  I 0
A=   .
Q I 0 λ I Q I
The geometric interpretation of this formula is as follows. We split the initial basis (e) =
(a , b )
in such a way that the number of vectors in a is the dimension of L := im(A − λ I) +
ker(A − λ I)m−1 . Obviously,   
 dim L + dim L = N . A new basis (aL , b ) in which A becomes
A P 
a block triangular matrix is
0 λ I
||L (a ) ||L (b )
(aL , b ) := (a , b )(ΠL (b ) + ΠL ).
We choose the part aL of the basis in such a way that its linear span L (aL ) contains the
image of A−λ I, hence the matrix of A−λ I has zero lower block in this basis. The construction
im(A − λ I) + ker(A − λ I)m−1 ensures the constancy of the Jordan normal form of A , namely,
it has the same Jordan form as A.

The construction gives a rational map from the orbit O to (C × C)ΣO , where ΣO is the total
number of all pairs P, Q. To prove that it is a birational isomorphism, it suffices to prove that
all functions P, Q are independent.
The Jordan form of A is determined by the numbers rank(A − λI)m . If these ranks have
the same values for all2 P, Q, the assertion will be proved.
The following fact is fundamental for the construction: the Jordan normal form of an
upper triangular matrix with scalar diagonal blocks is determined by these blocks only on an
algebraically open subset of the space of free matrix elements.
It is not difficult to see that our scheme gives a bijection between all Jordan normal forms
and the sets of diagonal blocks. To introduce the isomorphism explicitly, let us enumerate the
diagonal blocks of an upper triangular matrix according to their sizes nk :
λ In1 , λ In2 , λ In3 , . . . , λ Inmax −1 , λ Inmax , nk ≥ nk+1 ,
where det(A−λ I) = 0, nk = nk (λ ). The minimum size of a diagonal block corresponding to λ
def
is nmax × nmax , and we put nmax +1 = 0. The total number of diagonal blocks corresponding
to λ is denoted by max = max(λ ), and In is the identity n × n matrix.
The number of m × m Jordan blocks corresponding to λ = λ is equal to nm − nm+1 .
Now we are through with the An series and turn to the subject of the paper, which is the
series Bn , Cn , and Dn .

Let us split the initial basis (e) into three parts (a, c, b) and make two steps of the process,
one after another, the first one with the kernel of A − λ I and the second one with the image
2From an algebraically open set.

833
of A + λ I constructed using, in general, different eigenvalues λ and −λ . The result is
⎛ ⎞⎛  ⎞⎛ ⎞−1
I 0 0 λ I pgl pgl I 0 0
⎜ 1 2
⎟
A = ⎝q1gl I 0⎠ ⎝ 0 AI,I AI,II ⎠ ⎝q1 I 0⎠ ,
gl

q2gl 0 I 0 AII,I AII,II q2gl 0 I

  ≈
 −1
AI,I AI,II I 0 A pgl I 0
:= gl 3 ;
AII,I AII,II q3 I 0 −λ I q3gl I
consequently,
ω = ω ≈ + tr dpgl gl gl gl gl gl
1 ∧ dq1 + tr dp2 ∧ dq2 + tr dp3 ∧ dq3 .
J
≈ ≈
A fixed matrix similar to A is denoted by J .
Taking into account symmetries, one can convert the sum of the 2-forms into the canonical
expression. The calculation of the ranks (A − λ I)m shows that the constructed functions are
independent at least if ker A = ker A2 . Now we return to the main line of this paper.

2. The linear algebra of complex spaces with scalar products. Isotropic

subspaces
The groups of type B, C, and D preserve
 a nondegenerate
 bilinear form, a scalarproduct
. . . , . . . . The product is symmetric ξ, η = η, ξ for B, D, and antisymmetric ξ, η =
− η, ξ for C.
We consider all cases simultaneously, so we often use the word orthogonal in a broad sense,
for the symplectic scalar product too.
Consider one of the groups from the list and denote it by G; the corresponding algebra is
denoted by g. Let an index n run over the symmetric range n = ±[N/2], ±([N/2] − 1), . . . , ±1
and take the value n = 0 for odd N .
Definition 1. A basis is called standard if its Gram matrix3 g is
⎛ ⎞
0 0 τ  
⎝ 0 1 0⎠ , 0 τ 0 τ
, or ,
−τ 0 τ 0
τ 0 0
where τ is the square anti-diagonal matrix with ones on the anti-diagonal. It is the matrix of
the inversion: ⎛ ⎞
0 0 ... 0 1
⎜0 0 . . . 1 0⎟
⎜ ⎟
⎜ . .. .. ⎟
τ = ⎜ ... ..
. . . . .⎟ .
⎜ ⎟
⎝0 1 . . . 0 0⎠
1 0 ... 0 0
For all series B, C, and D, we have gT = g−1 . For the orthogonal groups g2 = I, for the
symplectic groups g2 = −I.

We treat g, τ as symbols of variable size, like the identity matrix I. This means that the
square matrices g, τ, I have sizes determined by the context, by the current formula in the text.
We consider linear algebra over C, so that there are nonzero isotropic vectors: ξ, ξ = 0
does not imply that ξ = 0.
Definition 2. A space L is called isotropic if it consists of isotropic vectors: ξ ∈ L ⇒ ξ, ξ = 0.
3The Gram matrix of a set of vectors f , f , . . . is the matrix of their pairwise products g := f , f .
1 2 ij i j

834
 If a standard basis is given, an example of an isotropic space is the coordinate subspace
spanned by several coordinate vectors with indices of the same sign.
The orthogonal complement L⊥ to a space L is the set of all vectors orthogonal to all vectors
of L:
η ∈ L⊥ ⇔ η, ξ = 0 for every ξ ∈ L.
The orthogonal complement is a subspace. For a nonzero isotropic space L, we have L ⊂
L⊥ = V , so that L + L⊥ = L⊥ = V . Nevertheless,
• (L⊥ )⊥ = L,
• dim L + dim L⊥ = dim V .
It is obvious that the dimension of an isotropic subspace does not exceed half the dimension
N of the space. Consequently, two isotropic subspaces in general position do not intersect.
Proposition 1. The set of pairs (E, G) of isotropic spaces of the same dimension n = dim E =
dim G is an algebraic manifold. It has an algebraically open subset such that
E ⊥ ⊕ G = G⊥ ⊕ E = V.
Let E ⊥ ⊕ G = V . Denote W = E ⊥ ∩ G⊥ . It is obvious that dim W = dim V − 2n, hence
V = E ⊕ W ⊕ G.
Proposition 2. The restriction of the scalar product . . . , . . . from V to W is nondegenerate
and has the same type as V .
Proof. Let ξ0 be such that ξ0 , η = 0 for every η ∈ W ; then ξ0 ∈ W ⊥ . By definition,
W := E ⊥ ∩ G⊥ , hence ξ0 ∈ E ⊥ , ξ0 ∈ G⊥ , and ξ0 ∈ (E + W + G)⊥ = V ⊥ = 0. 

Fix a splitting V = E ⊕ W ⊕ G, and let L be an isotropic space of the same dimension as

E. Let these spaces be from algebraically open sets such that L ⊕ W ⊕ G = L⊥ ⊕ G = V . Now
we define a special orthogonal linear transformation of V that transforms4 E into L. This
transformation consists of the following two projections: first we project E to L along W ⊕ G,
and then project L ⊕ W to L⊥ along G. The subspace G remains unchanged.

Let us denote by Q ∈ End V the transformation

||E ||W ⊕G ||L⊕W ||G
Q := (ΠW ⊕G + ΠL ) ◦ (ΠG + ΠL⊥ ).
The first transformation (from the left) is identical on W ⊕ G, it moves E to L parallel to
W ⊕ G. The second transformation is identical on G, it moves L ⊕ W to L⊥ parallel to  G.
⊥ ||L⊕W ||G 
Note that L ⊂ L , hence the second transformation is identical on L too: (ΠG + ΠL⊥ ) =
L
idL ∈ End L.
||E ||W ⊕G ||L⊕W ||G
Theorem 1. The transformation Q := (ΠW ⊕G + ΠL ) ◦ (ΠG + ΠL⊥ ) preserves the
scalar product, it is orthogonal and unimodular: Q ∈ G.
Proof. Let us introduce the following notation. Let a, c, b be sets of vectors. We write the
Gram matrix of the set (a, c, b) as a block matrix:
⎛ ⎞
a, a a, c a, b
⎝ c, a c, c c, b ⎠ ,
b, a b, c b, b
4It follows from the orthogonality that it also transforms E ⊥ into L⊥ .

835
where, for example, a, c is the matrix of pairwise products of vectors from the sets a and c,
the matrix element ( a, c )ij being equal to ai , cj .
Consider the transformation Q of the standard basis: (a, c, b) → (a, c, b)Q =: (aL , cL , b).
The corresponding matrix is block triangular, we denote its blocks by q, q , q:
⎛ ⎞
I 0 0
(a, c, b) ⎝ q I 0⎠ = (a + cq + bq , c + bq, b) = (aL , cL , b).
q q I
Our aim is to prove that the Gram matrix does not change. First of all, we note that L is
isotropic and aL belongs to L, so that aL , aL = 0.
The second set of basic vectors cL = c + bq extends aL to a basis of L⊥ ; consequently,
cL ⊂ L⊥ and aL , cL = 0.
Finally, the products of vectors from the set a + cq + bq with vectors from the set b are
the same as for the sets a and b, because L (c, b) = L (b)⊥ . Thus we have checked the first
line of the Gram matrix.
Consider the second line. The Gram matrix is symmetric or anti-symmetric, so we do not
need to check some entries, say c, a , b, a , and b, c . Consider c + bq, c + bq ; it keeps the
initial value g = c, c , because only vectors from the set b are being added to vectors from
the set c, but vectors from b are orthogonal to vectors both from b and c: L (b)⊥ = L (b, c),
whence c, b = 0 and b, b = 0. By the same reason, cL , b = c + bq, b = 0.
Formally, we must check one block b, b in the third line, but vectors from b have not been
changed at all. 

3. Symmetries of matrices from the group G and its algebra g

F
The group G  F changes the basis: (a, c, b) −→ (a, c, b)F, F ∈ G. It preserves the scalar
product, i.e., satisfies ξ, η = ξ T gη = (F ξ)T gF η, if and only if
F T gF = g.
Differentiating the equation F T gF = g gives the condition for A = Ḟ F −1 ∈ CN ×N to belong
to the algebra g:
A ∈ g ⇔ AT g + gA = 0.
The most important assertions are formulated as three theorems.
Theorem 2. For any vectors ξ, η the equation Aξ, η = − ξ, Aη holds.
Proof. We have
Aξ, η = (Aξ)T gη = ξ T AT gη = −ξ T gAη = − ξ, Aη . 
Theorem 3.
AT g + gAT = 0 ⇒ dim ker(A − λI)k = dim ker(A + λI)k for any k, λ.
Proof. We have
dim ker(A − λI)k = dim ker(AT − λI)k = dim ker(g−1 AT g − λI)k
= dim ker −(A + λI)k = dim ker(A + λI)k . 
We see that the eigenvalues of a matrix A ∈ g form pairs ±λ , and the structures of their
root spaces coincide. The third important fact is that the orthogonal complement to the
eigenspace corresponding to λ is the co-eigenspace corresponding to −λ .
Theorem 4.
ker⊥ (A − λ I) = im(A + λ I).
836
Proof. Let ξ = (A + λ I)ξ  and Aσ = λ σ; then (A + λ I)ξ  , σ = λ ξ  , σ + Aξ  , σ =
λ ξ  , σ − ξ  , Aσ = λ ξ  , σ − λ ξ  , σ = 0, whence
im(A + λ I) ⊂ ker(A − λ I)⊥ .
The structures of the root spaces of λ and −λ coincide, hence the dimensions of im(A + λ I)
and ker(A − λ I)⊥ coincide too, which implies that
im(A + λ I) = ker(A − λ I)⊥ , ker(A − λ I) = im(A + λ I)⊥ . 
Corollary 1.
(ker(A − λ I) ∩ im(A − λ I)k )⊥ = im(A + λ I) + ker(A + λ I)k .
Proof. It is obvious that (L ∩ M )⊥ ⊃ L⊥ + M ⊥ , and the dimensions coincide again. 

Proposition 3. The eigenspaces ker(A − λ I) and ker(A − λ I) are orthogonal if λ + λ = 0.
Proof. Let ξ  and ξ  be eigenvectors corresponding to λ and λ . Then
Aξ  , ξ  = λ ξ  , ξ  = −λ ξ  , ξ  ⇒ (λ + λ ) ξ  , ξ  = 0. 
Corollary 2. The eigenspaces corresponding to nonzero eigenvalues are isotropic.
Let us consider the zero eigenvalue. From the equation ker⊥ (A−λ I) = im(A+λ I) it follows
that ker⊥ A = im A. We have proved the following assertion.
Proposition 4. The kernel of A contains the isotropic subspace ker A ∩ im A.
Note that this subspace is zero if and only if there are no nontrivial Jordan blocks corre-
sponding to λ = 0.

Let us consider an arbitrary eigenvalue λ . All spaces ker(A−λ I)∩im(A−λ I)m are isotropic
for nonzero λ ; and if λ = 0, then the spaces are isotropic for m > 1. If λ = 0, from now on
we assume that m > 1.

Let us consider the transformation Q from Theorem 1. Let a splitting (a, c, b) correspond
to the dimension of the isotropic space L := ker(A − λ I) ∩ im(A − λ I)m determining the
transformation Q. This means that dim L (a) = dim L (b) = dim ker(A − λ I) ∩ im(A − λ I)m .
We consider the matrix Q−1 AQ of the transformation A in the basis (aL , cL , b), which is the
result of acting by Q = Q(λ , m) on the initial basis (a, c, b).
Theorem 5. The matrix Q−1 AQ ∈ g is block triangular:
⎛  ⎞
λI ρ ρ
Q−1 AQ = ⎝ 0 Aw ρ ⎠,
0 0 −λ I
where ρ, ρ , ρ are some matrices. The matrix Aw belongs to the algebra5 of the same type, i.e.,
Aw T g + gAw = 0, but is of smaller size.
Proof. The first set aL of basic vectors belongs to the eigenspace corresponding to λ , hence
the first column is (λ I, 0, 0)T .
By the definition of Q, the sets aL , cL form a basis of (ker(A − λ I) ∩ im(A − λ I)m )⊥ . By
Corollary 1,
(ker(A − λ I) ∩ im(A − λ I)m )⊥ = im(A + λ I) + ker(A + λ I)m ,
5The matrix (symbol) g is of “variable size,” here its size is by 2 dim ker(A − λ I) ∩ im(A − λ I)m > 0 smaller
than in the previous formulas.

837
hence the span of aL , cL contains the image of A + λ I, so that in the matrix of the transforma-
tion A+λ I in the basis aL , cL , b, which is Q−1 AQ+λ I, the lower dim ker(A−λ I)∩im(A−λ I)m
lines are zero.
The last thing we have to prove is that ATw g + gAw = 0. Since Q ∈ G, we have QAQ−1 ∈
g ⇔ (QAQ−1 )T g + gQAQ−1 = 0.
The situation is similar to the following one. The symmetry B T = B of any matrix B
implies the symmetry of any square subblock of B whose position is symmetric with respect
to the diagonal. Now we turn back to the algebra g. Since g is anti-diagonal, it follows that
B T g + gB = 0 implies Bw T g + gB = 0 for any square subblock B whose position is symmetric
w w
with respect to both the diagonal and the anti-diagonal. 

We reformulate the result taking into account the change of the Jordan form of any trans-
formation after taking its restriction to the subspace that contain the co-eigenspace and taking
the quotient by a subspace of the eigenspace (see [4]).
Let us consider the linear space
W := (im(A + λ I) + ker(A + λ I)m )/(ker(A − λ I) ∩ im(A − λ I)m ).
To obtain W , we take the quotient of the space of vectors orthogonal to the isotropic space L
by this space L. The result inherits the scalar product from V . It is not difficult to see that
the Gram matrix g of the standard basis of W has the same type as for V , but a smaller size.
A transformation A ∈ g ⊂ End V acts naturally on W ; we denote this action by Aw .
Theorem 6. The transformation Aw belongs to the algebra g of dim W × dim W matrices.
The Jordan normal form of Aw differs from the Jordan form of A by the number of blocks
corresponding to ±λ only. If λ = 0, the sizes of blocks of sizes m × m and larger become
smaller by one. If λ = 0, the sizes of blocks corresponding to the zero eigenvalue of sizes
m × m, m > 1, and larger become smaller by two.
The transformation from the basis (a, c, b) to the basis (aL , cL , b) is given by the block
triangular matrix Q: ⎛ ⎞
I 0 0
(a, c, b) ⎝ q I 0⎠ = (aL , cL , b).
q q I
It follows from Theorem 1 that the basis (aL , cL , b) is also standard. This implies that the
matrix Q belongs to the group G. Let us find out what this means for the blocks q, q, q and
for the blocks of Q−1 AQ.

4. Symmetries of block matrices

Let us introduce the operation of taking the conjugate matrix with respect to the anti-
diagonal: A → A . It transforms rows into columns and vice versa, but preserves the positions
of the elements lying not on the diagonal, but on the anti-diagonal:6
A−i,−j = A−j,−i ⇔ Ai,j = A−j,−i .
This anti-diagonal conjugation can be expressed via the inversion and the usual conjugation
A → AT :
A = (Aτ )T τ = τ AT τ,
where τ is the matrix of the inversion, which is the anti-diagonal matrix with ones on the
anti-diagonal.
6If the indices run from 1 to N , then A 
N−i+1,N−j+1 = AN−j+1,N−i+1 ⇔ Ai,j = AN−j+1,N−i+1 .

838
 To present the symplectic case, we need one more notion, namely, the operation that changes
the signs of entries with indices of one sign and preserves the entries with indices of the opposite
sign. The corresponding matrix is diagonal, with one half being the identity matrix I and the
other half being the minus identity matrix −I. In the case of orthogonal groups, we do not
need such an operation, and in a unified exposition we may set this matrix to be merely the
identity matrix.
To avoid extra notation, we note that the matrix gτ has all the necessary properties. In the
cases with a symmetric scalar product, it is the identity matrix, and in the symplectic cases, it
is the diagonal matrix with ones in the upper left half and minus ones in the lower right half.
Recall the following convention.
Remark 4.1. The matrices g, gτ , g2 are “adjustable,” like τ or the identity matrix. Their
structures are given, but the sizes depend on the context.

Now we turn to the important condition AT g + gA = 0, which is equivalent to A ∈ g. It can

be written as A = −τ gAτ g, where the involution A → τ gAτ g is identical in the orthogonal
cases.
Consider the symplectic cases. Let us write matrices in the block form with four blocks of
half the dimension. The involution preserves the diagonal blocks and changes the signs of the
anti-diagonal blocks:  
B C B −C
τg τg = .
E D −E D
Let us split a matrix A into 9 = 3 × 3 blocks in such a way that the pattern is symmetric
with respect to both diagonals. For the orthogonal groups, the condition AT g + gA = 0 is
equivalent to
⎛ ⎞
B ρ ρ
A = ⎝E F −ρ ⎠ , where ρ = −ρ , F  = −F, H  = −H.
H −E −B 


For the symplectic cases, τ g = I, and the condition AT g + gA = 0 is equivalent to

⎛ ⎞
B ρ ρ
A = ⎝E F −τ gρ ⎠ , ρ = ρ , F  = −τ gF τ g, H  = H.
H −E τ g −B 


We see that the middle block in the last column is determined by ρ, and we will write −τ gρ
instead of ρ.

Let us consider the lower unitriangular matrix Q. It follows from the symmetry QT gQ = g
that we can also replace the block q with −q  τ g:
⎛ ⎞
I 0 0
⎝q I 0⎠ .

q −q τ g I
The corner block q does not have a simple symmetry, it is neither symmetric nor antisym-
metric with respect to the anti-diagonal conjugation. Nevertheless, the matrix q has the
following property:
1
q = q − q  τ gq,
2
where q = ∓q has the same symmetry as ρ , and the term − 12 q  τ gq is a square matrix
of the opposite symmetry with respect to the anti-diagonal conjugation as compared with ρ
839
and q . The matrix q  τ gq is symmetric for the orthogonal groups and antisymmetric for the
symplectic groups, because (τ g) = ±τ g.
• For the orthogonal groups,
1 1  1 1
q := (q − q ), q τ gq = q  q = (q  q) .
2 2 2 2
• For the symplectic groups,
1 1  1
q := (q + q ), q τ gq = − (q  τ gq) .
2 2 2
We can write this in a uniform way:
     
q = q − (g2 )q /2, q  τ gq = (g2 ) q  τ gq ,
because g2 = ±I. We put g2 into brackets to emphasize that τ g and (g2 ) have different sizes
in the formula, this factor is merely the sign.

An arbitrary7 element from the orbit Og is represented as

⎛ ⎞⎛  ⎞⎛ ⎞−1
I 0 0 λI ρ ρ I 0 0
A=⎝q I 0⎠ ⎝ 0 Aw −τ gρ ⎠ ⎝∗ I 0⎠ .

q −q τ g I 0 0 −λ I ∗ ∗ I
5. Canonical coordinates on orbits of series B, C, and D
Let us regard a matrix A ∈ g as an element of the orbit of GL(N ) in gl(N ), where we know
Darboux coordinates (see [3, 4]). We represent the transformation
   
||L⊕W ||G ||E ||W ⊕G
Q := ΠG + ΠL⊥ ◦ ΠW ⊕G + ΠL
||E ||W ⊕G ||L⊕W ||G
from G as a pair of successive transformations ΠW ⊕G + ΠL and ΠG + ΠL⊥ , each of
them from SL(N ).
Let us denote the canonical coordinates corresponding to the first step by pgl gl gl gl
1 , q 1 , p2 , q 2 ,
and the coordinates corresponding to the second step by pgl gl
3 , q3 :
⎛ ⎞⎛  ⎞⎛ ⎞−1
I 0 0 λ I pgl pgl I 0 0
⎝q1gl I 0⎠ ⎜ ⎟
1 2
⎝ 0 AI,I AI,II ⎠ ⎝q1 I 0⎠ ,
gl
gl
q2 0 I 0 AII,I AII,II q2gl 0 I
    −1
AI,I AI,II I 0 Aw pgl 3
I 0
:= gl ,
AII,I AII,II q3 I 0 −λ I q3gl I
so that
ρ = pgl gl gl
1 + p2 q 3 , ρ = pgl
2, q = q1gl , q = q2gl , −τ gρ = pgl
3, −q  τ g = q3gl ,
and pgl 
1 = ρ + ρ q τ g. We can calculate the increment of the symplectic form when going from
the space V to the space W :

tr dpgl gl gl gl gl gl
1 ∧ dq1 + tr dp2 ∧ dq2 + tr dp3 ∧ dq3 = 2 tr dρ ∧ dq + tr dρ ∧ dq + tr dρ q τ g ∧ dq.
We use the equation
1
tr dρ ∧ (q + (g2 )q )/2 = tr dρ ∧ q  τ gq = 0
2
7An element from the algebraically open set of matrices with LU decomposition into the product of an upper
triangular and a lower triangular factors.
840
to replace q with its (anti)symmetric part q := (q − (g2 )q )/2. This is correct, because the
trace of the product of a symmetric and an antisymmetric matrices vanishes.
The matrix q = (q − (g2 )q )/2 is (anti)symmetric, and we can calculate the trace tr dρ ∧
dq using only half of the pairs of matrix elements ρ and q − (g2 )q = 2q ; these are exactly
the values that we need (along with q and ρ) to reconstruct the initial matrix. They form a
coordinate set of functions.
To formulate the main result of the present paper, now we introduce a pair of rectangular
matrices P and Q.
The matrix P consists of two blocks, one being rectangular and the other one being square.
The rectangular block 2ρ + ρ q  τ g is of size n × (N − 2n). The square block adjoined to the
right side of the rectangular one is ρ , of size n × n.
The matrix Q consists of the blocks symmetric to those of P . The upper block is q, of
size (N − 2n) × n. The square block adjoined to the lower side of the rectangular one is
q − (g2 )q − ð, where ð = adiagq is an anti-diagonal matrix8:

 
 q
P = 2ρ + ρ q τ g, ρ , Q = .
q − (g2 )q − ð
The version of these formulas for the orthogonal groups9 is

  q
P = 2ρ + ρ q  , ρ , Q = .
q − q
The version of these formulas for the symplectic groups is
  
 I 0 q
P = 2ρ + ρ q , ρ , Q = .
0 −I q + q − ð

Remark 5.1. About the anti-diagonal term ð. Let us illustrate the subtraction of the anti-
diagonal matrix ð. We consider the 2 × 2 case, the square blocks only, just for illustration:
  
a b d e 1 (h + f )/2 e
ρ = , q = , q = (q + q ) = .
c a g f 2 g (h + f )/2
The goal is to introduce variables p1 , q1 , p2 , q2 , p3 , q3 in such a way that the value tr dρ ∧ dq
takes the canonical form dp1 ∧ dq1 + dp2 ∧ dq2 + dp3 ∧ dq3 . A direct calculation gives
dρ ∧ dq = da ∧ d(h + f ) + db ∧ dg + dc ∧ de.
We have introduced the matrices
 
0 e h+f e
P = ρ , ð = adiag q = , Q = q + q − ð = ,
g 0 g h+f
and have claimed that the canonical coordinates are the pairs (P−1,−1 , Q−1,−1 ) = (a, h + f ),
(P−1,1 , Q1,−1 ) = (b, g), (P1,−1 , Q−1,1 ) = (c, e).
We have proved that any matrix A from an algebraically open subset of the orbit Og(J) can
be written as Qfin Pfin Q−1 k k k
fin , where Pfin = Pfin (Pij , Qij ) and Qfin = Qfin (Qij ) are the matrices
10

constructed during the iterative process. We number the steps of the process by k = 1, 2, . . . .
We have constructed a map Og → (C × C)ΣO  ∪k,i,j (Pijk , Qkij ), where ΣO is the number of
pairs of the constructed functions.
8The subtraction of ð divides the anti-diagonal of the lower square block of Q by two, in the symplectic case.
9We do not subtract ð, because in the orthogonal cases we do not use the anti-diagonal elements.
10The matrix Q consists just of Qk . Any matrix element of P is linear in the P k ’s, and the coefficients
fin ji fin ij
of these linear functions are products of matrix elements Qkji , linear in each Qkji .

841
Remark 5.2. Matrices from the algebraically closed sets may lie outside the orbit. As an
example, put Pijk = 0 for any i, j, k. We obtain diagonalizable matrices, which cannot belong
to a nondiagonalizable orbit.
Nevertheless, the Jordan form is constant on an algebraically open subset of (C × C)ΣO .
The image of the inverse map belongs to some orbit. Unfortunately, in some special cases
it is not the orbit we started from. Sometimes the functions constructed from the orbit are
not independent. Additional symmetries on Pijk , Qkij must be put on the construction of the
map the orbit. The loss of simplicity happens if the root space of the zero eigenvalue has a
complicated Jordan structure.
Theorem 7 (Concluding theorem). Canonical coordinates on an algebraically open domain
of the orbit Og(J), where dim ker J = dim ker J 2 , are the symmetric pairs of elements of the
matrices P and Q, namely, Pij and Qji , where
• the indices i, j for the orthogonal groups satisfy the inequality i + j < 0,
• the indices i, j for the symplectic groups satisfy the inequality i + j ≤ 0.

Proof. We have proved that the symplectic form takes the canonical form in the constructed
functions Pijk , Qkji . It remains to prove that for an algebraically open subset of the space of
parameters Pijk , Qkji , the formula A = Qfin Pfin Q−1
fin gives a matrix with the preassigned orbit.
It is not difficult to see that the Jordan form of any upper triangular matrix with fixed
scalar diagonal blocks is constant on an algebraically open subset of the space of free matrix
elements. Thus we have to prove only the independence of the functions Pijk , Qkji .
Let us apply the iterative process using only pairs of nonzero eigenvalues. At the end we
obtain a zero-dimensional orbit corresponding to the zero eigenvalue if it exists, this orbit does
not produce the functions Pij , Qji .
The independence of the constructed functions follows from the calculation of the Jordan
form of the upper triangular matrix Pfin during the iteration. To control the Jordan form, it
is useful to take into account the following.
• If matrices from the algebra g are similar to each other (i.e., have the same Jordan
structure), then they belong to the same orbit Og. This means that the conjugating
matrix can be taken from the corresponding group G.
• Matrices have the same Jordan form if and only if their functions κ of two variables λ
and m coincide: κ(λ, m|A) = rank(A − λI)m .
• The change of the Jordan form during the iteration for the eigenvalues ±λ = 0 is the
same as considered in [4]. 
5.1. Examples. In the first example, we demonstrate that the matrix elements of P, Q can
be dependent in the nilpotent case.
Example 1. The set of functions Pij , Qji constructed on the orbit O(J) where
⎛ ⎞
0 0 0 1 0
⎜0 0 0 0 −1⎟
⎜ ⎟
J =⎜⎜0 0 0 0 0⎟ ⎟ ∈ so(5)
⎝0 0 0 0 0⎠
0 0 0 0 0
is not independent. From the identity J 2 = 0 it follows that ρ = (ρ−2,0 , ρ−1,0 ) = 0, hence

(ρ )−2,1 q0,−1 (ρ )−2,1 ∗
P = ,
−(ρ )−2,1 q0,−2 ∗ ∗
842
 ⎛ ⎞
q0,−2 q0,−1
Q = ⎝(q )1,−2 − (q )2,−1 ∗ ⎠,
∗ ∗
so that the values Pij , Qji are dependent:
P−2,0 Q0,−2 = −P−1,0 Q0,−1 = P−2,1 Q0,−1 Q0,−2 .
Nevertheless, it is not difficult to construct a canonical coordinate set (pso so so so
1 , q1 ), (p2 , q2 )
using these Pij , Qji :
pso
1 = P−2,0 , q1so = 2Q0,−2 ; pso
2 = P−2,1 , q2so = Q−2,1 − Q0,−1 Q0,−2 .
We do not consider such cases in this paper. As a regular example of the theory, let us
calculate the canonical coordinates and check that they are indeed canonical in the simplest
case where the term ρ q  τ g does not vanish. This is the case n = 1, the nonzero matrices
ρ , q are 1 × 1, which implies the symmetry. The matrix g is an antisymmetric 2 × 2 matrix:
 
0 1 1 0
g= , gτ = .
−1 0 0 −1
We set Aw = 0, because the 2 × 2 case is not interesting, it coincides with the case of sl(2, C),
and any matrix B ∈ sl(2, C) satisfies B T g + gB = 0.
We parameterize the orbit Osp (J) of Sp(4, C) where J = diag(λ , 0, 0, −λ ) = 0 :
⎛ ⎞⎛  ⎞⎛ ⎞−1
1 0 0 0 λ ρ1 ρ2 ρ 1 0 0 0
⎜ q1 1 0 0⎟ ⎜ ρ2 ⎟ ⎜ 0 0⎟
A=⎜ ⎟⎜0 0 0 ⎟ ⎜ q1 1 ⎟ .
⎝ q2 0 1 0 ⎠ ⎝ 0 0 0 −ρ1 ⎠ ⎝ q2 0 1 0⎠
q q2 −q1 1 0 0 0 −λ q q2 −q1 1
The suggested coordinates on the orbit are
(psp sp sp
1 , p2 , p0 ) = (2ρ1 − ρ q2 , 2ρ2 + ρ q1 , ρ )
and (q1sp , q2sp , q0sp ) = (q1 , q2 , q ); below we will omit the superscript “sp” for the q-coordinates.
Note that the coordinates with subscript 0 come from the “anti-diagonal” of square 1 × 1
matrices, there is no sum of the form q + q here. In our formulas, the sum is nontrivial only
for non-anti-diagonal entries. The coordinates are just the matrix elements (Q−1 AQ)−2,2 and
(Q)2,−2 .

We immerse the orbit Osp (J)  A of the symplectic group to the “more roomy” orbit of
the general linear group Osl (J), where the canonical parametrization pi , qi , i = 0, 1, . . . , 4, is
known: ⎛ ⎞⎛  ⎞⎛ ⎞−1
1 0 0 0 λ p1 p2 p0 1 0 0 0
⎜q1 1 0 0⎟ ⎜ 0 ⎟ ⎜q1 1 0 0⎟
Av = ⎜⎝q2 0 1 0⎠ ⎝ 0
⎟⎜ ⎟⎜
⎠ ⎝q2 0 1 0⎠ ,
⎟
A
q0 0 0 1 0 q0 0 0 1
⎛ ⎞ ⎛ ⎞⎛ ⎞⎛ ⎞−1
1 0 0 0 0 p3 1 0 0
⎝ A ⎠ = ⎝ 0 1 0⎠ ⎝0 0 p4 ⎠ ⎝ 0 1 0⎠ .
q3 q4 1 0 0 −λ q3 q4 1
Equating the matrix elements, in particular, q0 = q , gives
⎛ ⎞⎛ ⎛ ⎞−1 ⎞
1 0 0 0 1 0 0 ⎛ ⎞−1
⎜ q1 1 0 0⎟ ⎜λ (ρ1 , ρ2 , ρ ) ⎝ 0 c
 1 0⎠ ⎟
A=⎜ ⎝ q2 0 1 0⎠ ⎝
⎟⎜ ⎟ ⎝∗⎠ ,
⎠
q2 −q1 1
q 0 0 1 0 A
843
 ⎛ ⎞⎛ ⎞ ⎛ ⎞−1 ⎛ ⎞⎛ ⎞ ⎛ ⎞−1
1 0 0 0 0 p3 1 0 0 0 0 ρ2
A = ⎝ 0 1 0⎠ ⎝0 0 p4 ⎠ ⎝∗⎠ = ⎝ 0 1 0⎠ ⎝0 0 −ρ1 ⎠ ⎝∗⎠ ,
q3 q4 1 0 0 −λ  q2 −q1 1 0 0 −λ
whence
1 sp 1
p1 = ρ1 − q2 ρ = (p − psp sp
0 q2 ), p2 ρ2 + q1 ρ = (psp + psp sp
0 q1 ),
2 1 2 2
1 1
p3 = ρ1 = ρ2 = (psp − psp sp
0 q1 ), p4 = ρ2 = −ρ1 = (psp + psp sp
0 q2 ),
2 2 2 1
q3 = q2 = q2sp , q1 = −q4 = q1sp , p0 = ρ = psp 0 , q0 = q = q0sp .
One can see that
4  2
dpi ∧ dqi = dpsp sp
i ∧ dqi .
i=0 j=0

Acknowledgment. I am grateful to Professor S. Yu. Slavyanov for his help and kind assis-
tance.

REFERENCES
1. I. M. Gelfand and M. A. Naimark, “Unitary representations of the classical groups,” Trudy
Mat. Inst. Steklov., 36, 1–284 (1950).
2. S. E. Derkachov and A. N. Manashov, “R-matrix and Baxter Q-operators for the noncom-
pact SL(N, C) invariant spin chain,” SIGMA, 2, 084 (2006); arXiv:nlin/0612003.
3. M. V. Babich and S. E. Derkachov, “On rational symplectic parametrization of the coad-
joint orbit of GL(N ). Diagonalizable case,” St. Petersburg Math. J., 22, No. 3, 347–357
(2011).
4. M. V. Babich, “Rational version of Archimedes symplectomorphysm and birational Dar-
boux coordinates on coadjoint orbit of GL(N, C),” Preprint MPIM Bonn, MPIM2010-59
(2010).
5. S. Lie (unter Mitwirkung von F. Engel), Theorie der Transformatiensgruppen, Abschn. 3,
Teubner, Leipzig (1893).
6. A. A. Kirillov, “Unitary representations of nilpotent Lie groups”, Russian Math. Surveys,
17, No. 4, 53–104 (1962).
7. V. I. Arnold, Mathematical Methods of Classical Mechanics, Springer-Verlag (1978).
8. V. Guillemin, E. Lerman, and S. Sternberg, Symplectic Fibrations and Multiplicity Dia-
grams, Cambridge Univ. Press, Cambridge (1996).
9. D. MacDuff and D. Salmon, Introduction to Symplectic Topology, Oxford Univ. Press,
Oxford (1998).
10. A. G. Reiman and M. A. Semenov-Tian-Shansky, Integrable Systems, A Group-Theoretic
Approach [in Russian], Institute of Computer Science, Moscow–Izhevsk (2003).

844