Physica 93A (1978) 327-353 © North-Holland Publishing Co.

DIAGONALIZATION OF THE QUADRATIC BOSON HAMILTONIAN

J.H.P. COLPA
Natuurkundig Laboratorium, Universiteit van Amsterdam, Nederland

Received 9 August 1977

Revised 30 January 1978

A general treatment has been given of the problem of diagonalizing a hamiltonian which is a
homogeneous quadratic expression in a finite number of boson construction operators. The
treatment provides us with several systematic procedures to diagonalize such hamiltonians in
practice; each algorithm in principle requires only a single unitary diagonalization of a hermitian
matrix.

1. Introduction

W e c o n s i d e r the diagonalization of a hamiltonian of the general f o r m (m

arbitrary positive integer)

H~. t
{a,'Dlr'rar t
Jr Ogr'O2r'rart+ aFD3,,~ar -F Otr'D4r'rOt*r}, (1.1)
r',r= 1

w h i c h is a h o m o g e n e o u s quadratic e x p r e s s i o n in a finite n u m b e r o f con-

struction o p e r a t o r s a,,ot*~. T h e s e o p e r a t o r s satisfy the w e l l - k n o w n c o m -
m u t a t i o n relations
[,~,,, * * - ~a~a~, = ~,,,,
Otr]-~ ~ Otr'Otr
[a,,, a~]-~ = [a;,, t
err]-, = 0, (1.2)

w h e r e r / = +1 or = - 1 d e p e n d i n g on w h e t h e r the m o p e r a t o r pairs are of the

b o s o n or f e r m i o n type, respectively. The coefficients in eq. (I.1) have b e e n
written as elements of the m - s q u a r e matrices /9,- (i = 1, 2, 3, 4) and can be
c o m p l e x n u m b e r s . T h e diagonalization of the hamiltonian (1.1) a m o u n t s to the
application of a h o m o g e n e o u s linear t r a n s f o r m a t i o n of the set of m pairs
a , , o t t, into a n o t h e r set of m c o n s t r u c t i o n - o p e r a t o r pairs ~/r,~/~, satisfying
c o m m u t a t i o n relations (1.2) with et replaced b y 3', w h i c h rewrites eq. (1.1) into
the f o r m

H = ~ hto,~l~,gr + constant. (1.3)

r=l

As is well k n o w n the s p e c t r u m o f H can be c h a r a c t e r i z e d just b y the values of

the m m o d e energies hto, w h i c h o c c u r in the right-hand side o f eq. (1.3).

327
328 J.H.P. C O L P A

A linear transformation of a set of construction operators into another such

set is called a Bogoliubov transformation after Bogoliubov ~) who carried out a
diagonalization by means of such a linear transformation. The same diagonal-
ization procedure had been used earlier by Holstein and Primakoff2). Some-
times, but not in this paper, the name Bogoliubov transformation is restricted
to a diagonalizing linear transformation.
In the fermion case the diagonalization of the hamiltonian is rather easy:
only unitary diagonalizations of hermitian matrices are involved. In this
connection we refer to the very neat and elegant method presented by Lieb et
al. 3) which is applicable in the case of real coefficients Dir'r. See also
subsection 6.4 for the more general case of complex coefficients; usually the
order of the matrices to be diagonalized is then two times larger.
In the boson case the situation is entirely different. Like in the fermion case
one can write down the "equations of m o t i o n " for the ~'r: [~/r, H] = hwr~/~
(r = 1,2 . . . . . m). After substituting for ~/r a linear expression in the OLand o(
with u n k n o w n coefficients and for H eq. (1.1), we easily derive using eq. (1.2)
an eigenvalue problem in ho~r:~ in which the eigenvectors give the coefficients
in the diagonalizing Bogoliubov transformation. At this point difficulties come
in because in general the problem does not appear as a unitary diagonalization
of a hermitian matrix so that one cannot use the theorems which exist for this
well-known problem. Many examples of diagonalizations have been worked
out in the literature but hardly any general a p p r o a c h can be found: no
systematic procedure is followed and usually the diagonalization takes place
in several steps, the whole procedure being very specific for the problem
under consideration§. Almost always the result is a hermitian, so-called
dynamical, matrix, the eigenvalues of which are the squares of the mode
energies. In accordance with this m a n y treatments seem to suggest 9) that the
dynamical matrix is the natural entity to look for when solving a boson
diagonalization problem. Although the knowledge of the dynamical matrix
m a y have great advantages for calculations so that it is worth looking for it, it
will be made plausible in this paper that its existence is not guaranteed. In
appendix A a theorem is proved which gives an impression of the strong
conditions on the coefficients in the hamiltonian (1.1) for the existence of the
dynamical matrix and which often can be applied for its construction.
Obviously a general a p p r o a c h to the boson diagonalization problem would
Essentially the same discussion can be given in the language of the Green-function formalism
leading to the same eigenvalue problem4). For the relatively simple problem under consideration
in this paper, use of Green functions does not seem to have any advantage. There exist other
ways than from the equations of motion to obtain the same eigenvalue problem: the lines of
a r g u m e n t which we shall follow in this paper to arrive at the eigenvalue problem is also very
familiarS).
§ C o m p a r e e.g. the diagonalizations in the theory of phonons~'7), in refs. 4 and 8 and in ref. 5.
 DIAGONALIZATION OF THE QUADRATIC BOSON HAMILTONIAN 329

be of great advantage. However, the most general approaches we could find

are contained in ref. 5 and in the book of Berezinl°). Lindghrd et al:) solve the
problem only for rather special types of hamiltonians, whereas Berezin does
not completely diagonalize the hamiltonian but indeed gives a method for
solving the problem~. This method, however, does not seem to be Very useful
in practice (see subsection 6.6 of this paper) and can only be applied to
hamiltonians with all coefficients real.
In the present paper we give an elementary but detailed matrix-algebraic
analysis of the whole problem of the diagonalization of a hamiltonian of the
form (1.1) in the boson case, i.e. under which conditions the diagonalization is
actually possible and how one can carry it out in practice. The algorithms
presented all involve only a single unitary diagonalization of a hermitian
matrix, for the computer solution of which problem techniques are available
which are much more advanced than those for the solution of a general
eigenvalue problem. In subsection 6.2 it will be indicated which of the
algorithms in an actual case is most efficient.
In sections 2-3 we treat the particular kind of quadratic hamiltonian which
occurs in solid-state physics; after that the most general hamiltonian (1.1) is
easily handled in section 4. A simpler diagonalization procedure, which can be
applied in many cases, is presented in section 5. We conclude this paper with
a discussion (section 6), including also the remark (subsection 6.4) that all
considerations of this paper are easily translated into the fermion case so that
for example also general fermion hamiltonians with complex coefficients can
be treated (cf. ref. 3). In appendix A a theorem concerning dynamical
matrices is proved and in appendix B we give some additional comments on
the matrix-algebraic terminology of this paper.

2. Diagonalization of the so-called Bogoliubov boson hamiltonian. Introduction

In this and the subsequent section we investigate the special case in which
the hamiltonian H has the form:

? ? *
Otr,A 2r,rOtm+r + OLm+r,A 3r,rOtr + ara+r ,A 4r'rOtm+r}, (2.1)
r',r= 1

where Ai (i = I, 2, 3,4) are m-square matrices with complex numbers as

,
elements. H contains m boson-operator pairs a~, ot*r and rn pairs am+r, ot~+,,

* Berezin was only interestedin the proof of the existence of a diagonalizingtransformationfor

the much harder case of an infinite number of constructionoperators.
330 J.H.P. C O L P A

i.e. 2m pairs in total, which satisfy commutation relations like eq. (1.2)
(~ = 1)~:
[O~o', ottp] = ~p,p, [Otp,, Otp] = * o~p]
]ap,, t = 0. (2.2)
Comparison of eq. ( 2 . 1 ) - i n v o l v i n g a set of 2m instead of m boson-operator
p a i r s - w i t h the more general class of hamiltonians (1.1), shows that the
hamiltonian (2.1) can be considered as a bilinear expression [terms like (a0) 2
where one operator is squared, do not occur] of a special type. A hamiltonian
of the form (2.1) is called a Bogoliubov hamiltonian. The reason why we
consider first eq. (2.1) is that it often occurs in practice [in solid-state physics
e.g. all operators with index r correspond to the same wave vector k, those
with m + r to - k ; m denotes the number of degrees of freedom in the unit
cell (or iess)]. Furthermore, having solved the diagonalization problem for eq.
(2.1), the diagonalization of the general form (I.1), though following a some-
what different scheme, is treated very easily (section 4).
The hermiticity of the operator (2.1) means that the 2m-square matrix§

~ ~-~ A3 A4 (2.3)

is hermitian. This matrix ~ plays a dominant role in the rest of the paper and
will be called the grand-dynamical matrix of the Bogoliubov hamiltonian (2.1).
We rewrite eq. (2.1) in matrix notation:

H---a*~a, ~t_~ (o~*.m+), ~---(\ ~ ~,

/
(2.4)
~rn+
where at and a are a row and column vector, respectively, each with 2m
construction operators as elements, and where a and am+ t are column m-
vectors.
In the introduction we indicated that the diagonalization of H amounts to a
homogeneous linear transformation of the set of construction operators
liltS, . . . . ~m,
f e~m+~
t . •.• , arm,
. . {~I,
. . ,. Otto,
. l~-m+l,
. . ~2,. into a suitable analogous
set of construction operators 3'. In section 3 we will prove that in the case of
eq. (2.1) we can choose a transformation

(3"t 3"rn+) ~ C* = ~t~-t __ (OC O/.m+)~-t and, equivalently,

(2.5)
3"m+ +

In this paper indices p ' , p can assume the values 1,2 . . . . . 2m and indices r', r the values
1,2 . . . . . m.
§ In this paper matrices often are represented in partitioned form: a 2m-square matrix is
represented by four m-square matrices and accordingly 2m-vectors by two m-vectors. Script
letters are used to indicate 2m-square matrices and 2m-vectors, Latin or Greek letters are used
for m-square matrices and m-vectors.
 DIAGONALIZATION OF THE QUADRATIC BOSON HAMILTONIAN 331

so that the 2 m - s q u a r e t r a n s f o r m a t i o n matrix 5 t r a n s f o r m s H into a diagonal

f o r m a c c o r d i n g to the s c h e m e

(2.6)

where ~ is a diagonal matrix

(5.*)-1~5.-1 = % ~ h d i a g ( o J 1 , o~2 . . . . . ¢o,., o , . + 1 , ~o.+2 . . . . . oJ2,.). (2.7)

The diagonalized hamiltonian is

2m
H =
r=l
(hmr'yt,~lr + ~(J)m+rVrtl+rVre+r)
* = E
p=l
t +
ho~,(3,o~', ½)
1 --
- ~ hip o hoJ,]. (2.8)
p=m+l

We c o n c l u d e this section by some i n t r o d u c t o r y r e m a r k s for the t r e a t m e n t in

section 3, where the conditions will be given u n d e r which the diagonalization
o f II is possible and w h e r e an algorithm will be p r e s e n t e d to p e r f o r m the
diagonalization in practiceS. Defining the " p a r a K r o n e c k e r s y m b o l " * g¢0 by

~o=I if l=<p_--<m,
gAo=-I if m+l_--<p_--<2m, (2.9)
8p,p = 0 if p' ¢ p,

we derive the conditions which should hold for the elements of 5. in order
that the ~, constitute a set of b o s o n operators p r o v i d e d the ,* do too:

~ ¢ p = [ ¢ ¢ , ¢ ~ ] = [ ( 5 . a ) ¢ , ( a * 5 . ' ) p ] = ~ 5.o'¢'5.*p"- ~ 5.¢p,,5.p*,,, (2.10a)

p"=l p"=m+l
or in matrix notation:
[Link]*=J or 5.*J5.=J or 5.,j=js.-1, ([Link])

where the " p a r a unit m a t r i x " J ~ - d i a g ( 1 , l . . . . . 1, - 1 , - 1 . . . . . - 1 ) (note:

,~p,p = 8p,p). H e r e the third and the s e c o n d e q u a t i o n are m o s t easily derived
f r o m the first and third equation, respectively. In view of the close r e s e m -
blance of eqs. (2.10a) and (2.10b) to the definition of unitarity we shall call a

:[:In case one is only interested in the prescriptions of the algorithm (not in the argumentation),
see section 3, remark 1, which can be read at this moment if one does not bother about the
meaning of the concept "para-unitary" and if one defines the 2m-square matrix J~ according to:
,~ m diag(l, 1. . . . . l, -1, -1 . . . . . -1).
* In the treatment of the fermion case on this place the Kronecker symbol 8¢p is encountered.
In situations like this we use the word or prefix "para" to provide a comparison with the fermion
treatment (see also subsection 6.4).
332 J.H.P. C O L P A

matrix para-unitary:~ if its elements satisfy one of the f o u r equivalent equa-

tions (2.10a) and (2.10b). Finally, b y taking the inverse of the first eq. (2.10b),
we see f r o m the s e c o n d definition eq. (2.10b) that the inverse ~ ~ of a
p a r a - u n i t a r y matrix 9- is also para-unitary.
N o t e that for the f e r m i o n case we w o u l d have o b t a i n e d instead of eqs.
(2.10a) and (2.10b) the unitarity condition for the matrix ~-. In view of this the
p r o b l e m of diagonalizing a f e r m i o n B o g o l i u b o v hamiltonian can be c o n s i d e r e d
as solved c o m p l e t e l y (see s u b s e c t i o n 6.4). T o see m o r e clearly the specific
difficulties related to the diagonalization p r o b l e m (2.7) in the b o s o n case, we
rewrite eq. (2.7) as:

(~-t) 1 ~ - - i = % __ h diag(wl, 092. . . . . to,,, to,,+,, o9,,+2. . . . . w~,,,)

---diag(Xb A2. . . . . A,,, --Am+b --A,,+2. . . . . --Z2,,) =-- , ~ ,
(2.11)

w h e r e the diagonal matrix 5g has been i n t r o d u c e d to simplify the notation.

Multiplying this e q u a t i o n on the left by 3 -+ we obtain a c c o r d i n g to the third eq.
(2.10b):

~-l = ~ - + ~ = ,~--t~. (2.12)

F r o m eq. (2.12) we see that the pth c o l u m n t~p--(u o vo)' (p = 1,2 . . . . . 2 m ; u o
and vo are c o l u m n m - v e c t o r s and the prime d e n o t e s a c o l u m n vector) of ~ t
satisfies

~r,~ = Xo.~rco or (~- Ao.~)rr'o = 0 or (a,

A3 a2"~(uo~
a4J\vo/ = Ao \-roe
( up "~, (2.13)

and for the diagonal e l e m e n t s a o of ~ we obtain the equation§:

det(~-a0})=0 or a , -A3a~/ a , +/12

Afl I =0. (2.14)

in which I is the m - s q u a r e unit matrix. The difficulties c o n n e c t e d with the

diagonalization problem (2.7) are n o w easily r e c o g n i z e d by multiplying eq.
(2.13) on the left by J :

(~-ao~)rr o=-(.~-ao~)rr, o=0 or \ -a3 -A a ~ #'o=O, (2.15)

¢ In the existing literature other prefixes are used for this concept, e.g. pseudo-unitary. In the
present treatment we use the prefix to indicate counterpart concepts so that it appears in many
words: para-eigenvector, paranorm, etc. but not in others: the counterpart (matrix) of a hermitian
matrix is also hermitian, n o t para-hermitian. To avoid unwanted interference with other concepts
we use the in matrix algebra unusual prefix "para-".
§ In this or almost this form the boson diagonalization problem is often formulated in the
literature4,~).
 DIAGONALIZATION OF THE QUADRATIC BOSON HAMILTONIAN 333

where # is the 2m-square unit matrix. The main questions to be solved

can now be formulated. (i) Are the eigenvalues of the (in general non-
hermitian) matrix ~ -- ~ real? (ii) Does a para-unitary matrix ~--1 exist with
the eigenvectors of ~ as columns? (iii) How can a para-unitary matrix ~--~
be calculated in practice which diagonalizes the matrix 5~ according to eq.
(2.7)? In section 3 answers to these questions are given and a practical
diagonalization algorithm is presented which involves only a single unitary
diagonalization of a hermitian matrix.

3. Diagonalization of the Bogoliubov boson hamiltonian. Conditions and

algorithm

In this section we investigate the conditions under which the diagonaliza-

tion of the boson Bogoliubov hamiltonian (2.1) is possible according to the
scheme (2.6). Since a hamiltonian has a lowest energy level the mode energies
htop in eq. (2.7) should be non-negative [cf. eq. (2.8)]. Now it will turn out (see
the last two examples in remark 4 at the end of this section) that mathematical
difficulties occur in case we allow zero to be a mode energy (soft mode).
Therefore we exclude this case from our treatment and give only a brief
discussion on this restriction in subsection 6.5. The remaining problem is to
find the conditions under which the para-unitary diagonalization of ~ accord-
ing to eq. (2.7) is possible with all diagonal elements in % positive. The answer
is given by the following theorem.

Theorem. A 2m-square hermitian matrix 5~ can be para-unitarily diagonal-

ized into a matrix with all diagonal elements positive if and only if ~ is
positive definite.
Proof of the necessity of the positive definiteness. Since (ff*)-~ = (~'-~)*, eq.
(2.7) has the form of a conjunctivity transformation [i.e. the form S*PS = Q, P
(and Q) being hermitian matrices and S a non-singular matrix]. For such
transformations we have Sylvester's law of inertia which states that all
conjunctivity diagonalizations of a certain (hermitian) matrix P lead to
(diagonal) matrices for which the numbers p, n and z of positive, negative and
zero diagonal elements, respectively, depend only on the matrix P and not on
the particular choice of the diagonalizing matrix S (the diagonal elements are
real because the matrix Q is hermitian). Since according to our assumptions
eq. (2.7) requires that the matrix ~ can be diagonalized by a conjunctivity
transformation into a matrix with all diagonal elements positive, we conclude
that the familiar unitary diagonalization of 5~ too should lead to a matrix with
all diagonal elements positive. This means that the matrix ~ should be
positive definite.
334 J.H.P. COLPA

Proof of the sufficiency of the positive definiteness. We prove this by

presenting a construction for the diagonalizing para-unitary matrix ~--1 in eq.
(2.7).
Since ~ is positive definite it is possible (see for example remark 1 after the
proof) to find a (non-singular) 2m-square matrix ~ which satisfies ~ = ~;*~K.
Let ~--~ be a para-unitary matrix which diagonalizes 5~ according to eq. (2.7);
we don't know yet whether such a matrix exists. If we denote in what follows by
%_+in the matrix with diagonal elements ep÷~/2 in case % is a diagonal matrix
with all diagonal elements ep > O, then the matrix °-It defined by
O~ ~ ~ - - - 1 ~ - - | / 2 (3.1)

is unitary since (# is the 2m-square unit matrix)

~ , ~ = ~-,/2(~--,)*~K*~K~--~~ ,/~ = ~ ,~2(~--%*~--~%-,n = ~ , / 2 ~ - , 2 = ~.
(3.2)

Moreover ~ diagonalizes the (hermitian) matrix .°K#~K*,i.e.

= %,n~-~-t%,n = %,/2~%1/2= 27, (3.3)

where the first eq. (2.10b) has been used and where the diagonal matrix 27 has
been defined by eq. (2.11).
Let us conversely choose a unitary matrix ~t which diagonalizes the
(hermitian) matrix ~K~* according to eq. (3.3) in such a way that the first half
of the diagonal of 2? contains only positive elements and the second half only
negative elements [Sylvester's law of inertia, applied to the first and last
member of eq. (3.3), tells us that the number of positive elements and also the
number of negative elements in 2? and # is the same]. Such a matrix
provides us by eq. (3.1) with a para-unitary matrix ~--' which diagonalizes
according to eq. (2.7). Indeed, since by eqs. (3.1) and (3.3):
(~-l)t#~- 1 : ~l/20~*(~t)-I~J~-10~l/2 : ~11227-1~1/2 = ~ , (3.4)

the matrix ~--t is para-unitary according to the second eq. (2.10b). Further-
more, $--J diagonalizes ~:
(~--i)*~- , = %,/:0lt*(~Kt)-,(3~*3~)~K-t~%l/:= %. (3.5)

With the existence of a matrix ~-~ with these properties we have shown the
possibility of the para-unitary diagonalization of ~ and the theorem has been
proved completely.

Remark 1. Algorithm. The proof given above suggests an algorithm

- e s s e n t i a l l y amounting to a single unitary diagonalization of a hermitian
 DIAGONALIZATION OF THE QUADRATIC BOSON HAMILTONIAN 335

m a t r i x - f o r finding a para-unitary matrix if-J, which diagonalizes a given

positive-definite hermitian matrix 5~ according to eq. (2.7). Firstly we have to
find a decomposition of the 2m-square positive-definite hermitian matrix ~ in
the form 5~ = ~t~f, where ~K is a 2m-square matrix. In computer programs it is
convenient to apply the so-called Cholesky decomposition (see for example
ref. 11, where the considerations for real matrices are easily generalized to
complex matrices) and find a (non-singular) complex matrix ~ which is
upper-triangular (i.e. an element vanishes if the column index is smaller than
the row index). The failure of the algorithm (by which we also understand the
case that only a singular matrix ~ can be found, see also subsection 6.5) in
case 5~ is not positive definite can be used as a check on this property of ~ .
The computer time needed for a Cholesky decomposition is entirely negligible
in comparison with the time entailed in a unitary diagonalization of a hermi-
tian matrix, which also can be used to find a decomposition (see the end of
subsection 6.5). Secondly we diagonalize unitarily the hermitian matrix ~K#~K*
and find a unitary matrix °-tt, in which the columns are arranged in such a way
that [see eq. (3.3)] the first m diagonal elements of the diagonalized matrix
~t(~Jff£*)~t = ~ are positive, the last m negative; the diagonal matrix ~ in eq.
(2.7) is then given by g = J.~. Finally we calculate the matrix if-1 row by row,
beginning with the last row, from eq. (3.1), written in the form ~ / 2 = ~ i f - I .
For testing computer programs it may be useful to realize that 2m-square
para-unitary matrices f are easily constructed from arbitrary m-square uni-
tary matrices T. For example, if T is such a matrix and if c and s are any
complex numbers with Icl 2- Is Is= 1, then the matrix

~.=_(cT s*T*~
\sT c'T*] (3.6)

is para-unitary (see the first footnote in section 4 for the meaning of T*). We
further remark that no additional calculation of if is needed once if-~ is
known. For with the help of the last eq. (2.10b) we see that for a para-unitary
matrix if holds:
-- V t

Remark 2. The generalized eigenvalue problem. Eq. (2.13) has the form of
what is known in matrix algebra as the "generalized eigenvalue problem":
(P - AQ)w = 0, (3.8)
where P and Q are n-square matrices; the complex constant A and the n-
vector w are called a generalized eigenva|ue and a corresponding generalized
eigenvector, respectively. In general the way of solving this problem (i.e.
336 J.H.P. COLPA

determining values for A and w) depends on the properties of P and Q. Our

p r o o f of the sufficiency of the positive definiteness a b o v e is based on a
well-known solution of the problem (3.8) in the case that P and Q are
hermitian matrices and m o r e o v e r P is positive definite. One just p e r f o r m s the
decomposition P = KtK and finds the matrix R with Q = K t R K . Eq. (3.8) can
then be written

K*(A ~I-R)Kw=-O or (R-A ~I)Kw=O, (3.9)

which obviously is the simple eigenvalue problem connected with the hermi-
tian matrix R (eigenvalues h -~, eigenvectors Kw).

Remark 3. Another point of view. In the proof a b o v e it was not necessary

for the construction of the matrix ~- ' to consider its individual columns [cf.
question (ii) at the end of section 2]. In m a n y cases [for example in case m = 1
so that the problem (2.13) can be solved by hand (see examples in remark 4
below), in section 4 and in appendix A] another point of view is convenient.
For p ' and p fixed eq. (2.10a) expresses a relation which holds between the
elements of the p ' t h and pth row in a para-unitary matrix. Expressing the
second eq. (2.10b) in terms of the matrix elements we see that similar
relations hold b e t w e e n the columns. N o w we call two 2m-vectors 1~- (Ul vl)'
and t~,z----(u2 v2)' (u and v are column m - v e c t o r s and the prime denotes a
column vector) paraperpendicular (to each other) iftr,rt~t¢,-~ u*iR 2 - I;IV2= * 0
and denote the real value * ~ - = u ' u - v t v as the p a r a n o r m of a vector
rr'-~ (u v)' (note that one cannot change the sign of a p a r a n o r m by multi-
plication of the vector by a scalar). Then we can characterize a 2m-square
para-unitary matrix as a matrix in which the columns (or rows) are paraper-
pendicular and paranormalized, i.e. the p a r a n o r m of each of the first m
columns equals 1, that of each of the last m columns - 1 . Calling f u r t h e r m o r e
the vectors •'0 in eq. (2.13) para-eigenvectors of ~ at the para-eigenvalue h o (in
what follows short: p a r a v e c t o r s and paravalue), a c o m p a r i s o n with eq. (2.1 I)
shows that the problem in section 2 was to find 2m paravalues of 9 , the first
m of which should be positive, the last m negative. In order to find the
para-unitary matrix ~-~ we need 2m p a r a v e c t o r s of ~ which correspond to
the paravalues and which are paraperpendicular. The positive definiteness of
ensures that the first (last) m of them have a positive (negative) paranorm
and after (para)normalization the p a r a v e c t o r s can serve as the first (last) m
columns of ~--~. For we have for a paravectorr¢,=- (u v)' of ~ at paravalue h :
 DIAGONALIZATION OF THE QUADRATIC BOSON HAMILTONIAN 337

so that the paravalues of such ~ are real (this was seen already in the proof of
the theorem above) and positive and negative paravalues correspond to
paravectors with positive and negative paranorms, respectively. Note finally
that paravectors ::,~ and :::2 of a positive-definite hermitian matrix ~ at different
paravalues A~ and A2 are paraperpendicular, which is easily seen by subtrac-
ting from

t Ul
(3.11)
_

the complex conjugate with all subscripts I and 2 interchanged.

Remark 4. Counterexamples. We have seen that a large class of hermitian

bilinear expressions of the form (2.1) can be diagonalized, viz. all expressions
with a positive- (or of course negative-)definite hermitian grand-dynamical
matrix (2.3). It is evident that this class does not exhaust all boson Bogoliubov
operators which can be diagonalized. A trivial example is the expression
H---a~oq-cqa~ which has the form (2.1) (with m = 1) and obviously can be
diagonalized [i,e. written into the form (1.3)], whereas the grand-dynamical
matrix ~ ~ diag(1,-1) is not positive (or negative) definite. However, not all
hermitian operators of the form (2.1) can be diagonalized (note that all ht%
should be real because in eq. (2.7) together with ~ also % is hermitian). For
example, it is not possible to diagonalize the expression (again m = 1) H ---
ot*~q + V'2(et~ot~ + a2oL~) + ot2ot2*,since the grand-dynamical matrix [A~ = A4 =
(1), A2 = A3 = (~/2) has the non-real paravalues i and - i (see for "para"
concepts used here and in what follows remark 3 above). Note that in this
example the paranorms [1 + il2 - I-~/2[ 2 and I1- il2 - I-V'2[ 2 of both paravec-
tors (1 + i -V'2)' and ( 1 - i -~/2)' vanish whereas for diagonalizability of the
expression (2.1) paranorms of +1 and - 1 are required (see remark 3). Even in
cases where the grand-dynamical matrix is positive semidefinite, the diagonal-
ization is not always possible. The expressions H = c2o~oq + CS(Ot~a*2+ oL2oq)+
s2ot2ot~ (c and s any real numbers with c 2 - s 2= 1) and H~ot~oq+
(a~ot~+~t2oL~)+a2o~ are examples of Bogoliubov operators with positive-
semidefinite grand-dynamical matrices, the first of which can be diagonalized
[into ~/]~,~(+0~,2)] whereas the second cannot since the corresponding grand-
dynamical matrix has only one paravector (1 -1)'. From the examples given
above we consider the restriction to positive-definite grand-dynamical
matrices quite natural from a mathematical point of view and satisfactory
from a physical point of view (see subsection 6.5 for a calculation in
connection with positive-semidefinite grand-dynamical matrices).
338 J.H.P. C O L P A

4. Diagonalization of the general quadratic boson hamiltonian

We now turn to the diagonalization of the general quadratic hamiltonian

(1.1):

H =- ~ {ar,D~r,,.~r + a , D z , , r ~ r + ~X~,D3~,ra, + a~,D4r,r~ tr}, (4.1)

r',r= l

where the m boson-operator pairs a~,a*~ satisfy the commutation relations

(1.2) (with 7/= 1). An operator of the form (4.1) is hermitian if and only if the
following relations hold:

D I r ' , + D4~, = D*rr'+ D~r,r, Dlrr = D*rr,

(4.2)
D2r, r + D2rr, = D~rr, + D~,~, l<=r',r<~m

(note: the first line of this equation implies that the diagonal elements of D~ and
D4 are real). For the subsequent analysis it is essential that

Dlr'r = D*rr, = D4rr,, D2r, r = D2,~, = D~,,.,,

1 ~ r', r _-< m, (4.3)
D3r, r = D 3 w = D~.,, D4r,r = D~rr, = Dlrr,,

which obviously can be assumed without restricting the generality: a possibly

necessary rewriting in this sense modifies the expression ( 4 . 1 ) - w h i c h is
supposed to represent a hamiltonian and thus a hermitian o p e r a t o r - a t most
by an additive (real) constant. Then the 2m-square matrix*

has the form shown in eq. (4.4) and is hermitian (which implies that A is
hermitian and B symmetric). We call ~ the grand-dynamical matrix of the
hamiltonian (4.1) and use it to rewrite eq. (4.1) in matrix notation:

at (4.5)

[cf. eq. (2.4); no confusion is to be expected that no extra symbols indicate

that a* and ,~ in a* are row m-vectors and in a column m-vectors]. To
emphasize that the meaning of the concept grand-dynamical matrix depends
on whether it belongs to a Bogoliubov hamiltonian of the form (2.1) or to a
general hamiltonian of the form (4.1), we rewrite the Bogoliubov hamiltonian

:[: By A* we denote the complex conjugate of the matrix A [elements: ( A * ) , , , = ( A r , r)*].

 D I A G O N A L I Z A T I O N OF T H E Q U A D R A T I C BOSON H A M I L T O N I A N 339

(2.1) in the form (4.1) where in eq. (4.1) m is replaced by 2m. Then the
hamiltonian (2.1), which we denote by Hsog for the moment, becomes in
matrix notation:

A I - Tr A4), (4.6)
HB°g~(a*am+act'+)½ 0 A~ A* 0
' /\ A , O o
l!
in which the square matrix, including the factor ½, is the grand-dynamical
matrix of the general quadratic hamiltonian which matrix by definition should
have the form of the right-hand side of eq. (4.4). The additive constant in eq.
(4.6) is easily expressed in terms of the mode energies ht% [cf. eq. (2.8)]:
2m

TrAI-TrA4 = h~oo - ~ hoJo. (4.7)

o=l p=m+l

One readily verifies, indeed, that the left-hand side of eq. (4.7) is an invariant
of @Bos, by which we denote the matrix in eq. (2.3) for the moment, under
para-unitary transformations; the right-hand side of eq. (4.7) is the value of
Tr A~- Tr A4 for the diagonal matrix in the right-hand side of eq. (2.7).
We now proceed with the diagonalization of the general quadratic hamil-
tonian (4.5). Taking into account that the matrix 3- in a transformation to new
construction operators ~,:

¢:3-, or (~,):(QP, O a
P*)(ot+) (4.8,

has the partitioned square form shown in eq. (4.8) and that just like in eq.
(2.10a) we can prove the para-unitarity of 3-, we have according to eq. (3.7):

,,.), (4.9)

Now in a diagonalization according to the scheme~

H =-a+5~a= at3-t(~"t)ll~)3--13-a = ¢tc/g~, (4.10)

with a matrix 3- of the form (4.9), the diagonal matrix ~ necessarily has the
form:

(3-+)-1~3--1 = % _ ½h diag(tob w2. . . . . to,., tol, 02 . . . . . tom). (4.11)

:~ Note that, although eq. (4.10) has completely the same form as eq. (2.6), the two schemes are
different: unlike the ot and ate,++ in eq. (2.5), the two m-vectors ot and ot+ in a [eq. (4.5)] essentially
contain the same construction operators.
340 J.H.P. COLPA

The special form in the right-hand side, because the matrix (,~*) 1~?~---I is
hermitian (so the diagonal elements are real), while its left-upper and right-
lower m-square matrices are each other's complex conjugate if ~ and ~ 1
have the partitioned forms (4.4) and (4.9), respectively.
After tbese introductory remarks it is very easy to show that again the
necessary and sufficient condition that a diagonalization according to the
scheme (4.10) with all w positive is possible, is that the grand-dynamical
matrix ~ is positive definite*. The necessity follows from eq. (4.11), in a
similar way as in section 3. The sufficiency is seen from the following
construction, also suitable for actual calculations, of a ~ t with the required
properties. Let us choose a matrix ~-~ which diagonalizes ~ para-unitarily.
According to the theorem in section 3 such a matrix exists but the one chosen
possibly has not the form (4.9). By comparing eqs. (4.11) and (2.11), we see
from eq. (2.13) that the first m columns (Ur Vr)' of ~-~ satisfy:
Aur + Bvr = ~l w , u r Av* + Bu* = -~hwrv*
or (4.12)
B*u,+A*v~=-~hwrvr B*v*+A*u*=~hwm*,l r = 1,2 . . . . . m,
so that (v* u*)' is a paravector at the paravalue -~hw~ if (u~ vr)' is a paravector
at the paravalue ½hwr (see for " p a r a " concepts section 3, remark 3). Since
obviously together with the (urvA' also the ( v ' u * ) ' ( r = 1,2 . . . . . m ) are
paraperpendicular to each other and m o r e o v e r - a c c o r d i n g to the discussion
around eq. (3.11) (different p a r a v a l u e s - t h e (v* u*)' are paraperpendicular
to the m vectors (ur vA', the matrix arising from the replacement of the last m
columns of the chosen ~--i by (v* u*)', (v* u~')', . . . . (v* u*)', is a para-unitary
matrix with the required properties: it has the form (4.9) and the diagonaliza-
tion property (4.11).
To summarize, we have shown that the general quadratic boson hamiltonian
(4.1) can be diagonalized with all mode energies positive if and only if the
grand-dynamical matrix (4.4) is positive definite. The diagonalized hamiltonian
is obtained by substitution of eqs. (4.11) and (4.8) into eq. (4.10):

H = z..,, h~or(V/y~+~). (4.13)

r=]

5. An often applicable simpler procedure for the diagonalization of quadratic

boson hamiltonians

In sections 3 and 4 the possibility has been proved of the diagonalization of

hamiltonians of the form (2.1) and (4.1). respectively. The proofs immediately
:~The arrangement of the elements in the matrix c~ [eq. (4.4)] is essential. For example,
Berezin~°) makes a different choice and consequently has to require the positive definiteness of
two matrices instead of one (see also subsection 6.6).
 DIAGONALIZATION OF THE QUADRATIC BOSON HAMILTONIAN 341

indicated the methods to p e r f o r m the diagonalizations in practice. Although

these methods were completely general, in special cases simpler procedures
m a y be preferable. Lieb et al. 3) described a procedure for diagonalizing a
general quadratic hamiltonian of the form (4.1) in the case that the et are
fermion operators and the coefficients Dir'r are r e a l . Like most of the consi-
derations in the fermion case (see subsection 6.4) also this procedure can be
translated into the boson case. This will be shown in this section because the
case of real coefficients often occurs in practice and because the single
unitary diagonalization to be p e r f o r m e d in the new procedure concerns a
hermitian matrix of order m (cf. the order 2m of the matrices in sections 2-4).
Starting f r o m a complex positive-definite hermitian 2m-square matrix ~ of
the form

we describe a practical procedure to find a para-unitary matrix ff-~ of the

form [cf. eq. (3.7)]:

U U* - V*

which diagonalizes ~ according to (in this section h~, A2. . . . . Am are always
positive):

(~*)-1~--1 = ~ ~_~ diag(Ab A2. . . . . Am, /~1, '~2 . . . . . Am). (5.3)

By definition [see the second eq. (2.10b)] the para-unitarity of ~--~ means that
[u, and Vr denote the columns of the m-square matrices U and V in eq. (5.2)]

U*r,l~r--1) tr,Dr:~r,r,
r', r = 1,2 . . . . . m. (5.4)
U¢r'l)r- IJtr'Ur : 0,

From eqs. (4.4), (4.9) and (4.11) we see that the procedure to be described
can be used for treating the general problem of section 4 in case of a real
grand-dynamical matrix ~:~ and according to eqs. (2.3) and (2.7) also in the
special cases of the problem of section 2 in which ~ , though possibly
complex, has the form (5.1). In the problem of section 4 the diagonalization
procedure follows the scheme (4.10) with a and ¢ defined by eqs. (4.5) and
(4.8), respectively; in the problem of section 2 the scheme (2.6) with a and ¢
defined by eqs. (2.4) and (2.5), respectively.

$ Lieb et al. restricted their considerations to this case (for fermions). Like all the considera-
tions in this paper also this section allows a self-evident translation into the real variant. Thus to
a real ~ of the form (5.1) one can find, along the lines to be described in this section, a real ~--1 of
the form (5.2) and consequently of the required form (4.9).
342 J.H.P. COLPA

By c o m p a r i n g eqs. (5.3) and (2.11), we see f r o m eq. (2.13) that the first m
c o l u m n s (Ur Vr)' and the last m c o l u m n s (vr u,)' of ~- 1 should satisfy:

A]\vr] "~(Uv,)and (A A]\ur]B'~(Vr'~ __,r(__Vur), (5.5a)

respectively. E a c h of the two equalities in eq. (5.5a) can be written m o r e

explicitly as

A u , + Bvr = A~Ur,
(5.5b)
Bur + Avr = -A~v~.

A d d i n g and subtracting the two equalities in eq. (5.5b) give the relations:

( A + B)rkr = ¢~r~Jr, (5.6a)

( A - B)q'r = A,ek~, Ckr -~ u~ + v,, qJr =- Ur -- Yr. (5.6b)

Multiplying eqs. (5.6) on the left by A - B and A + B, respectively, we obtain:

( a - B ) ( A + B)rkr = Ar(a - B)q,r = A 2~br, (5.7a)

( a + B ) ( A - B)0r = A,(A + B)&r = A 2Abr. (5.7b)

N o w the matrices ( A - B ) ( A + B) and (A + B ) ( A - B ) in general are not

hermitian (if so, it is possible and o f t e n preferable to apply the t h e o r e m in
a p p e n d i x A, see r e m a r k 2 at the end of a p p e n d i x A: both matrices then equal
A 2 - B 2, w h i c h matrix then is called the d y n a m i c a l matrix of the c o r r e s p o n d -
ing hamiltonian). T h e r e f o r e we p r o c e e d as follows. T h e hermiticity of the
matrix ~ implies that the matrices A and B and c o n s e q u e n t l y also A + B and
A-B are hermitian. F u r t h e r m o r e , t o g e t h e r with ~ also the matrices A + B
and A - B are positive definite. This is easily r e c o g n i z e d if one m a k e s for an
arbitrary c o l u m n m - v e c t o r u the evaluation: (u* +-u*)~(u +-u)' = 2u*(A +- B)u,
and realizes that the single e l e m e n t in this 1-square matrix is positive (see also
r e m a r k 2 below). W e n o w c h o o s e an m - s q u a r e matrix K with A - B = K ' K .
T h e n the eigenvalue p r o b l e m (5.7a) can be written as a unitary-diagonalization
p r o b l e m of a hermitian matrix:

K * [ K ( A + B ) K * - A 2fll(K*)-~&~ = 0
[ K ( A + B ) K * - A~llxr, X~ =- (K*) '&~. (5.8)

Starting f r o m an arbitrarily c h o s e n o r t h o g o n a l set of m e i g e n v e c t o r s X~ of the

(hermitian) matrix K ( A + B ) K * w h i c h are normalized in such a w a y that
X*rXr = A; ~ [take the Ar positive; together with A + B also K ( A + B ) K * is
positive definite[, we d e t e r m i n e the c o r r e s p o n d i n g ~b~ f r o m eq. (5.8) and G
f r o m eq. (5.6a) (always taking the A, positive); eq. (5.6b) then holds a u t o m a -
 DIAGONALIZATION OF THE QUADRATIC BOSON HAMILTONIAN 343

tically. N o w w e can s h o w that the ur and v, as derived f r o m these d,r and ~br
a c c o r d i n g to eq. (5.6), f o r which o b v i o u s l y eqs. (5.5b) and, equivalently, (5.5a)
holdS, r e p r e s e n t the c o l u m n s of the m a t r i c e s U and V in a p a r a - u n i t a r y
matrix ff-~ of the f o r m (5.2), which diagonalizes ~ a c c o r d i n g to eq. (5.3). In
fact, eq. (5.4) is satisfied b e c a u s e f r o m eqs. (5.7a), (5.6b) and (5.8) and in view
of our n o r m a l i z a t i o n of the Xr we have:

a 2Ab<qbr
t
= ~b*r,(A- B ) ( A + B)~b, = A,,cb~,(A + B)cb~
= A~,X*r,K(A + B)K+x~ = A2,8<~, (5.9)
so that
~b~,4~ = 8r, r or rather ( u ~ , - v*,,)(u~ + v,) = 8,,r,
(5.10)
~b*etk, = 8<~ or rather (u*~,+ v+,,)(u,- Vr) = 8r'r.

T h e two equalities in eq. (5.4) follow f r o m eq. (5.10) by adding and s u b t r a c -

ting. O u r p r o b l e m has b e e n solved.

R e m a r k 1. Like in r e m a r k 1 of section 3, in c o m p u t e r p r o g r a m s it is
r e c o m m e n d e d to c a r r y out a C h o l e s k y d e c o m p o s i t i o n to find a matrix K with
A - B = K+K. T h e p r o c e d u r e is very efficient and s i m u l t a n e o u s l y the positive
definiteness of A - B can be c h e c k e d without p e r f o r m i n g a unitary diagonal-
ization, which would take a lot m o r e c o m p u t e r time.

R e m a r k 2. In case one s u c c e e d s in the C h o l e s k y or a n o t h e r d e c o m p o s i t i o n

of A - B (i.e. A - B is positive definite) and if m o r e o v e r all eigenvalues of the
matrix K ( A + B ) K + are positive (i.e. A + B is positive definite), it is e n s u r e d
that ~ is positive definite also: no additional c h e c k is n e c e s s a r y to v e r i f y that
the c o n s i d e r a t i o n s of this section are applicable. T o recognize the positive
definiteness of 5~ we only h a v e to c o n s i d e r f o r an arbitrary n o n - z e r o c o l u m n
2 m - v e c t o r e~ =- (u v)' the single e l e m e n t in the I - s q u a r e matrix

~ ~ = ~ [(u ++ v +)(A + B)(u + v ) + (u +- v +)(A - B)(u - v )], (5.11)

which e v i d e n t l y is positive since the c o l u m n m - v e c t o r s u + v and u - v c a n n o t

vanish simultaneously.

6. Concluding remarks

6.1. Connection between the treatments o f the Bogoliubov and the general
quadratic hamiltonian

T h e f o r m (4.6) of the B o g o l i u b o v hamiltonian (2.1) contains a 4 m - s q u a r e

matrix (including the f a c t o r -~), which has the f o r m (4.4) required to call it the
~tIn the terminology of section 3, remark 3, the vectors (u, v,)' and (O r /dr)' are paravectors of
at the paravalues Ar and -At, respectively.
344 J.H.P. COLPA

grand-dynamical matrix of a general quadratic hamiltonian. Thus one may in

theory diagonalize a hamiltonian of the form (2.1) by the treatment of section
4 rather than by the treatment of sections 2 and 3. That eq. (4.6) never helps
to solve the diagonalization problem in the case that the treatment of sections
2 and 3 does not apply, follows from the fact that the two grand-dynamical
matrices -the 2m-square matrix (2.3) and the 4m-square matrix occurring in
eq. (4.6) -are both positive definite as soon as one of them is.

6.2. Algorithm to be used

The question which of the algorithms presented in this paper is to be used

when a homogeneous quadratic boson hamiltonian is actually to be diagonal-
ized, is easily answered. First of all one tries to write the hamiltonian,
possibly by adding a real constant, into the standard form (2.1) [the grand-
dynamical matrix (2.3) is then hermitian]; if this is not possible, into the
standard form (4.1) with a hermitian grand-dynamical matrix of the form (4.4).
If the resulting grand-dynamical matrix 9 has the form (5.1), the considera-
tions of section 5 can be used for the diagonalization. Otherwise one has to
use an algorithm which involves a unitary diagonalization of a hermitian matrix
of two times larger order: the algorithm described in section 3, remark 1 for the
standard form (2.1) [scheme according to eq. (2.6)] and the algorithm
described in section 4 for the standard form (4.1) [scheme (4.10)]. The
algorithm of section 4 follows completely the one in section 3, remark 1, but
only requires the calculation of the first rn columns of 3-l whereas the last m
columns must be obtained from the first according to eq. (4.9). A
simplification can be achieved if one can use the theorem in appendix A
(dynamical matrix, see subsection 6.3); therefore it is recommended always to
check whether that theorem or a theorem of that kind is applicable. All
considerations of this paper can obviously be translated into their real
variants, which implies another simplification for hamiltonians with all
coefficients real. Complications can arise in the case of soft modes (zero mode
energies, positive-semidefinite %a), see subsection 6.5.

6.3. Dynamical matrices

The approach in this paper differs widely from those used in almost all
special cases treated up to now in the literature. In solid-state physics, e.g.
one usually finds positive-definite hermitian matrices (called dynamical
matrices, see also section I), like the grand-dynamical matrices directly
constructed from the coefficients in the hamiltonian, the eigenvalues of which
yield the squares of the mode energies. In appendix A we discuss a connec-
 DIAGONALIZATION OF THE QUADRATIC I]OSON I-IAMILTONIAN 345

tion between our results and those in the literature: it is made plausible that
rather than being a general result the occurrence of dynamical matrices is
specific for a very limited class of quadratic boson hamiltonians.

6.4. The fermion counterpart

All considerations in this paper, including those in appendix A, can in a

self-evident way and without any difficulties be translated into the fermion
case. The " p a r a " language of this paper then passes into the familiar matrix-
algebraic language: the diagonalization of a homogeneous quadratic fermion
hamiltonian amounts to the unitary diagonalization of a grand-dynamical
matrix ~ which is constructed from the coefficients in exactly the same way
as in the boson case. Also in the fermion case 5~ is hermitian; no additional
properties like positive definiteness are required, however, for the diagonal-
izability of the hamiltonian. A decomposition like Cholesky's (cf. section 3,
remark 1) is not involved. Since results of this kind for the fermion case are
well known (see also section 1), we do not consider this case further. The
discussion in appendix B on some more details of the " p a r a " terminology of
this paper gives further help to see the diagonalization problem in the boson
and fermion case as counterpart problems.

6.5. Soft modes (zero mode energies)

The most unsatisfactory aspect of this paper is that we did not include in
our treatment the case of a quadratic boson hamiltonian with a positive-
semidefinite grand-dynamical matrix ~ , which often occurs in practice. In the
last two examples in remark 4 at the end of section 3 we saw that not all
positive-semidefinite matrices can be para-unitarily diagonalized. In ac-
cordance with this it is essential that the matrix ~ is positive definite in order
that the algorithm in section 3, remark 1 for finding a diagonalizing para-
unitary matrix ~--1 can be applied. However, we want to point out that there
exists a simple calculation in connection with a positive-semidefinite matrix
which often is physically interesting. To be specific we turn our attention to
solid-state physics where a positive-semidefinite grand-dynamical matrix
often occurs for certain isolated values k0 of the wave vector k (often k0 = 0).
If one does not bother about the possible non-existence of a diagonalizing
para-unitary matrix ~-~ whose elements may tend to infinity, the algorithm in
section 3, remark 1 can be used to calculate from the positive-semidefinite
the (limit of the) mode energies for k tending to such a k0. T o see this we
recognize - e.g. by considering a positive-semidefinite matrix as a limiting case
of positive-definite matrices - from section 3, remark 1 that the unitary diagonal-
346 J.H.P. COLPA

ization of the hermitian matrix ~ E ~ + gives the desired m o d e energies. Rather

than trying to generalize the procedure of the Cholesky decomposition to the
case of positive-semidefinite matrices ~ , one can apply for this single case an
additional unitary diagonalization to find a decomposition 5~ = ~t3~: if 5~ is a
unitary matrix with 5e+5~5~ = Y diagonal, one easily verifies that we can take
~ = ~ + = 5e~'/zb°+, where the meaning of the positive-semidefinite diagonal
matrix 9 '/: is self-evident. Analogous arguments can be used in section 5 to
find the Ar (not the 9--~) also in case of positive-semidefinite matrix (5.1).

6.6. Comparison with other approaches

In section 1 we already indicated in what aspects our treatment differs from

those in the literature. In most publications we referred to, special (solid-
state) problems were solved. It m a y be interesting to c o m p a r e our a p p r o a c h
with a purely theoretical treatment. As such the book of Berezin m) seems to
be most suitable: theorem 8.1 of ref. 10, to which we referred already in
section 1 and in the last footnote of section 4, is concerned with almost the
same problem as section 4 of this paper. H o w e v e r , the contents of both
a p p r o a c h e s are by no means identical.
On the one hand Berezin includes the case of an infinite n u m b e r of
construction operators. This means a much more difficult problem to solve
and simultaneously makes the mathematics less accessible. On the other hand
Berezin treats only the case of real coefficients and does not diagonalize
completely: in his final expressions terms with ~/*r'~'rOccur where r' ~ r so that
an additional (unitary, see below) transformation is necessary for a complete
diagonalization. Also no explicit algorithm is given in ref. 10. Expressed in the
notation of section 4 with real matrices A and B the treatment of ref. 10 just
results in the real matrix:
(R'/2+ R t/2 R , / ~ _ R ,,,2) R ~ P ,/2(p,/2Qp~/2),/2p ,.,~,
~-t=~ I/2_R ,/2 RE/2+R ,/2, P-~A+B, Q=-A-B. (6.1)

This matrix has the required partitioned form (4.9) and one indeed v e r i f e s
directly (see below) that this ~ ~ transforms the matrix ~ [eq. (4.4)] para-
unitarily into a 2m-square matrix (3-+) ~ 5 - ~ with the two off-diagonal m-
square blocks zero. H e r e a f t e r the diagonalization of the diagonal blocks can
easily be p e r f o r m e d by an additional unitary diagonalization, that is by a
para-unitary and at the same time unitary matrix which has the form of the
right-hand side of eq. (3.6) in which T is an m-square orthogonal (i.e. real and
unitary) matrix and c = l+ s = 0. The "square r o o t s " in eq. (6.1) are defined
according to the example of P: p~/2 is the real positive-definite symmetric
matrix for which P ~/2p,/2 = p, and P ~/2=_(p i/2) ~ (see section 5 for the proof
 DIAGONALIZATION OF THE QUADRATIC BOSON HAMILTONIAN 347

that ~ is positive definite if and only if the same holds for P ~ A + B and
Q = A - B). Indeed, the matrix (6.1) is para-unitary since it satisfies the first
eq. (2.10b). Moreover the off-diagonal blocks mentioned vanish according to
the simple reduction
[R InPR in _ R -lnQR-in] = ~R - I n [ R P R - Q]R-~/2 = 0. (6.2)

Since in general one needs to carry out a unitary diagonalization for each
"square root" occurring in eq. (6.1) (see the end of subsection 6.5), one
recognizes from eq. (6.1) that Berezin did not intend to give a practical
algorithm for the calculation of a para-unitary matrix with the properties
mentioned, but just wanted to present a proof for its existence: our treatment
(cf. section 5 which is applicable in the problem as considered by Berezin)
requires only a single unitary diagonalization of a hermitian matrix of the
same order as the matrices which are to be unitarily diagonalized in the
procedure suggested by ref. 10.

Appendix A

The d y n a m i c a l matrix

In section t and subsection 6.3 we announced a discussion on the

frequently encountered dynamical matrices. By a dynamical matrix we mean
a hermitian matrix D, the eigenvatues of which give the squares of the mode
energies. According to this paper these mode energies can also be obtained
from the paravalues of a grand-dynamical matrix ~ ; in general the order of D
is half of the order of ~ . In this sense the matrices ( A - B ) ( A + B ) and
( A + B ) ( A - B ) occurring in eq. (5.7) are obviously only dynamical matrices if
they are hermitian (see also remark 2 at the end of this appendix). In this
appendix we prove an almost trivial theorem which we believe to be a typical
example of the kind of theorems which lie behind the existence of dynamical
matrices. Application of the theorem yields a matrix D and allows to find the
coefficients in a corresponding diagonalizing Bogoliubov transformation [i.e.
the elements of the matrix ~-~ occurring in eqs. (2.7) and (4.11)] from the
eigenvectors of D. Since a matrix decomposition like Cholesky's (cf. section
3, remark 1 and section 5, remark l) is not involved, the relations between the
elements of :ff~ and the coefficients in the original hamiltonian are more direct
than if for example the is found with the help of the considerations in
section 5; this may have advantages in actual calculations. Although the
requirements which the theorem imposes on the elements of ~ are rather
strong, a check on a great many investigations in the solid-state literature
348 J.H.P. COLPA

shows that, though not in all cases where a dynamical matrix exists, in the
vast majority of such cases one could have derived the dynamical-matrix
results also by application of just this theorem. This makes the theorem an
almost systematic key for finding dynamical matrices D (in case of their
existence) but simultaneously makes it plausible that the existence of the
dynamical matrix in many theories is indeed only due to the peculiar relations
which hold between the coefficients in the hamiltonian and that the dynamical
matrix in general is not naturally connected with the boson diagonalization
problem as the grand-dynamical matrix is.

Theorem. Let A and B be commuting (complex) m-square matrices; A is

hermitian, B is normal (i.e. BB* = B i B , e.g. B is hermitian). Let furthermore the
m-square hermitian matrices A and D -- A 2 - B * B be positive definite. Then the
2m-square hermitian matrix

is positive definite also and its paravalues and paravectors (and also the
diagonalizing para-unitary matrix) can be constructed from the eigenvalues
and those sets of m orthogonal eigenvectors of D where these eigenvectors
are simultaneous eigenvectors of the matrices A and B. Each p-fold (positive)
eigenvalue of D is the square of p positive and p negative paravalues of the
matrix ~ .
Proof. On the conditions of the theorem the (hermitian) matrices in the
left-hand side of the equation

(A, BA)( AB, -BA)=(A2-B*Bo A:-OB*B) (A.2,

commute. Consequently there exists an orthonormal set of 2m simultaneous

eigenvectors of these two matrices. Let (u v)' be an eigenvector of this set
with eigenvalues x, and x2, respectively. We have

A u + By = x~u, A u - By = x2u,
(A.3)
B*u + A v = xjv, - B * u + A v = x2v,

and by adding the upper and the lower two equations we obtain:
2 A u = (xj + x2)u and 2 A v = (xt + x2)v, (A.4)

respectively. Since A is positive definite and since the vectors u and v do not
simultaneously vanish, x~ + x2 is positive. On the other hand (u v)' is also an
eigenvector of the matrix in the right-hand side of eq. (A.2) at the product xtx~,
as eigenvalue which is positive since A 2 - B*B is positive definite. So both xt
 DIAGONALIZATION OF THE QUADRATIC BOSON HAMILTONIAN 349

and x 2 are positive and we see that all eigenvalues x~ of ~ [~ is the first matrix
in the left-hand side of eq. (A.2)] are positive: ~ is positive definite.
Since B is normal, B is unitarily diagonalizable. Because A and B commute
and A is hermitian one can even find a unitary matrix which diagonalizes A
and B simultaneously. Since B is normal such a unitary matrix also diagonal-
izes the matrix B* and consequently the matrix D =- A 2 - B*B. In what follows
we start from the columns ~r (r = 1, 2 . . . . . m) of such a unitary matrix or
from another orthogonal set of m simultaneous eigenvectors of A and B (and
automatically also of B* and D). The corresponding eigenvalues of D are
denoted by A ~ (in what follows we take always Ar > 0).
We presently prove that the 2m 2m-vectors constructed from the m-vectors
~r according to:

r-,[(Aq-ArI)~r~ (Ur) ¢.., [ - B ~ r ~ (Um+r'} (A.5)

-~ ' [l~m+r~ ~m+rk(A + ArI)~r] ~ \Vm+r/

constitute a set of paraperpendicular paravectors of ~ at paravalues Ar and

-X,, respectively [in eq. (A.5) the scalars Cr, Cm+, are determined such that
the vectors are ("para")normalized: U'rUt-- Vt,Vr = 1 and IA,m+rUm+r --
l) •ra+rl)ra+r -1]. According to section 3, remark 3, these paravectors
consequently can serve as the columns of a para-unitary matrix

~ - - l ~_~ (g//~ ¢'~'2 • • • f~'m f~'m+l t~'m+2 • • • f ~ ' 2 m ) , (A.6)

which diagonalizes ~ , i.e.

(~-,)-t~--i = ~ --- diag(Ai, A2. . . . . Am, AI, A2. . . . . Am) (A.7)

[cf. eqs. (2.7) and (4.11)].

It is easily checked that the vectors (A.5) are paravectors of ~ at the
paravalues mentioned, indeed: we have

( BA B'~[(A + t~rI)~r~ = A~/(A + $t~rI)~r~

A}k -B*~ ] k B /~r ]
and (A.8)
a)( a
B
+ xA)~r/ (A + xA)~d"

The first m vectors (A.5) and also the last m are paraperpendicular: U~'Ur--
v~,v, = 0 and Um+rU' m+rt' __ l)m+r,l)m+ = 0 (r' ~ r), because Ur, 1)r, Um+r and l)ra+r all
equal 6r multiplied by a scalar. Moreover the first m vectors (A.5) are
paraperpendicular to the last m since they correspond to different (because of
the different sign) paravalues [see the discussion around eq. (3.11)]. We
conclude that the 2m vectors (A.5) are paraperpendicular. With this we have
proved the theorem completely.
350 J.H.P. COLPA

Remark 1. We constructed the 2m paravectors (A.5) of ~ from a particular

orthogonal set of m eigenvectors of D =-A 2 - B'B, viz. from an arbitrary set
of simultaneous eigenvectors of the matrices A and B. That such a particular
choice is necessary can be seen from the following counterexample (m = 2):

A-=(~ ]), B = - ( 22 ~), D=-A 2-B 2=(~ 0). (A.9)

We have

A ( 1 ) = 5 ( I ), A(/1)=3(_ll), B(1)=4(]), B(ll) =0(_11). (A.10)

According to our construction we should start from eigenvectors sc~ and ~2

proportional to (1 1)' and (1 - I ) ' , respectively, and then obtain after a simple
calculation which follows eqs. (A.5) and (A.6) the para-unitary matrix

x/---g (i: ) 0

0
2

2 -xG
'
(A.I1)

which diagonalizes ~ according to eq. (A.7) with % ----diag(3, 3, 3, 3). However,

if we would take the set of orthogonal eigenvectors of A 2 - B 2, where
s¢1~- (1 0)' and so2~- (0, 1)', we would get according to eq. (A.5) (AI = A2 = ~/9 =
3) the column vectors (not normalized):

(A.12)

Obviously these vectors do not form a paraperependicular system.

Remark 2. It is clear that the theorem is applicable to the problem of

section 5 if there the matrices ( A - B ) ( A + B) and (A + B ) ( A - B) are hermi-
tian. The ~ in this appendix has then the form (5.1) so that the matrices A and
B are hermitian. ~ is positive definite, so also A. The hermiticity of (one of)
the products ( A - B ) ( A + B) and (A + B ) ( A - B ) implies that the matrices A
and B commute and that both products equal D =- A 2 - B 2. From eqs. (5.7) we
see that the matrix D is positive definite.
 DIAGONALIZATION OF THE QUADRATIC BOSON HAMILTONIAN 351

Appendix B

On the "para" terminology of this paper

In several places in this paper we introduced "para" concepts (cf. section 2
and section 3, remark 3), the properties of which strongly suggest that the
diagonalization problem of homogeneous quadratic hamiltonians in the boson
and fermion case is governed by counterpart matrix-algebraic formalisms (cf.
also subsection 6.4). The fermion case need not be considered any further
since the connected concepts are well known. Therefore we only give a few
details concerning the boson case: some points of view which were not used
in this paper, may contribute to a better understanding of the "para" concepts
used.
The central "para" concept is the para unit matrix § defined by:

~ ( I 0 0 i ) - = - d i a g ( 1 , 1 . . . . . 1 , - 1 , - I . . . . . -1), I unitmatrix, (B.1)

where here and in what follows it is understood that the order of script-letter
matrices and vectors is 2m and the order of Latin-letter matrices and vectors
m. Next we define the para-inverse 3 of an arbitrary non-singular matrix ff by

if---( RP Q ) ' ~-'=(~" ~)' ~---(?O -S/~) (B.2a)

(the second equation defines the m-square matrices /5, Q, /~ and S), or,
equivalently, by one of the following equivalent three definitions:
~J~=J or ~J~-=J or ~'---,~5-tJ. (n.2b)
From the first two of these equations we see that each non-singular matrix is
the para-inverse of its para-inverse: 3-= ~'. Furthermore, since (~'-')-' = ~',
the last definition equation (B.2b) implies that

i.e. the para-inverse of the inverse of a (non-singular) matrix is the inverse of

its para-inverse.
Now we can place some considerations of this paper in a new .ligtit. From
eqs. (2.10b) and (B.2b) we see that a 2m-square matrix ~ is para-unitary if and
only if
3* = ~. (B.4)
[Using that one can express the para-inverse of a product of matrices in a
reversed-order product of para-inverses of the individual matrices (this is
352 J.H.P. COLPA

seen f r o m the reduction ~ - ~ J g ° ~ = ~-J~-=.~), one easily verifies from eq.

(B.4) that the product of a n u m b e r of para-unitary matrices is again para-
unitary.] Consequently eqs. (2.7) and (4.1 I) can be written in the form [cf. also
eq. (2.11)]:

--- diag(Xl, A2
. . . . . A,,, --Am+l, --Am+2 . . . . . -A2,,) --= j ~ , (B.5)
where we replaced ~--i by 9° and where ~ also is given in partitioned form.
We recognize that this equation has not the form of a similarity trans-
formation (i.e. the form S-~PS = Q; as is well k n o w n such a transformation
does not affect the eigenvalues) we are a c c u s t o m e d to. Instead, it has the form
of a parasimilarity transformation (by which we mean the f o r m ~ 9 p = ~),
which type of transformation does not affect the paravalues as we will show
now. In a g r e e m e n t with section 3, r e m a r k 3 we call a non-zero column vector
t~, = (u v)' a p a r a v e c t o r of an arbitrary matrix 5~ (in what follows we do not
require the hermiticity of ~b) at the paravalue A if

~e,J=AJ or (5~-AJ)r,,:0 or (C B)(u)=A(Uv) (B.6,

[cf. eq. (2.13)]. F r o m eq. (B.6) we see immediately that the paravalues A of a
matrix ~ are given by the roots of the " p a r a s e c u l a r equation":

det(~-A~)=0 or IA-AI
C D +BA I I =0 (B.7)

[cf. eq. (2.14)]. By considering the determinant of the matrix ~ b g ~ - A ~ =

~ ( ~ b - A.~)9p [see the first eq. (B.2b); we do not require here the para-unitary
of 9°: ~¢ should only be non-singular in order that it has a para-inverse], we see
that a parasimilarity transformation does not affect the paravalues, indeed.

Acknowledgements

We are very m u c h indebted to Dr. A.H. 't Hooft, Dr. P.E. B r o m m e r and
Dr. H.G.J. Pijls for several valuable discussions and suggestions which
appreciably contributed to this paper. We are also grateful to the referee for
useful r e m a r k s on a previous version of this paper. We thank Dr. B r o m m e r
for critically reading the manuscript.

Note added in proof

After submission of this p a p e r to P h y s i c a another p a p e r on the same

subject, by C. Tsallis, appeared in the Journal of Mathematical Physics 19
 DIAGONALIZATION OF THE QUADRATIC BOSON HAMILTONIAN 353

(1978) 277. It may be interesting to compare the two treatments which differ in
many respects.

References

1) N. Bogoliubov, J. Phys. (USSR) 11 (1947) 23, also reprinted in D. Pines, The Many-Body
Problem (W.A. Benjamin, Inc., New York, 1961) p. 292.
2) T. Holstein and H. Primakoff, Phys. Rev. 58 (1940) 1098.
3) E. Lieb, T. Schultz and D. Mattis, Ann. Phys. 16 (1961) 407, appendix A.
4) C.F. Coil, III and A.B. Harris, Phys. Rev. B4 (1971) 2781.
5) P.A. [Link], A. Kowalska and P. Laut, J. Phys. & Chem. Solids 28 (1967) 1357.
6) A.A. Maradudin, E.W. MontroU, G.H. Weiss and I.P. Ipatova, Theory of Lattice Dynamics in
the Harmonic Approximation, 2nd edn. (Academic Press, New York, London, 1971).
Solid State Physics, H. Ehrenreich, F. Seitz and D. TurnbuU, eds., Suppl. 3, section 11.3.
7) R.D. Mattuck, A guide to Feynman Diagrams in the Many-Body Problem (McGraw-Hill Publ.
Co. Ltd., London etc., 1967) section 16.2.
8) C.F. Coil, III and A.B. Harris, Phys. Rev. B2 (1970) 1176.
9) R.J. Elliott, J.A. Krumhansl and P.L. Leath, Rev. Mod. Phys. 46 (1974) 465, section II.A.2.
10) F.A. Berezin, The Method of Second Quantization (Academic Press, New York, London,
1966) theorem 8.1.
11) J.H. Wilkinson, The Algebraic Eigenvalue Problem (Clarendon Press, Oxford, 1965) p. 229.