On the mathematical treatment of the

Born-Oppenheimer approximation.
Thierry Jecko

To cite this version:

Thierry Jecko. On the mathematical treatment of the Born-Oppenheimer approximation..
2013. <hal-00803621v3>

HAL Id: hal-00803621

https://hal.archives-ouvertes.fr/hal-00803621v3
Submitted on 4 Jul 2013 (v3), last revised 23 Jan 2014 (v5)

HAL is a multi-disciplinary open access

archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.

Larchive ouverte pluridisciplinaire HAL, est

destinee au depot et `a la diffusion de documents
scientifiques de niveau recherche, publies ou non,
emanant des etablissements denseignement et de
recherche francais ou etrangers, des laboratoires
publics ou prives.

On the mathematical treatment of the

Born-Oppenheimer approximation.
Thierry Jecko
AGM, UMR 8088 du CNRS, Universite de Cergy-Pontoise,
Departement de mathematiques, site de Saint Martin,
2 avenue Adolphe Chauvin, F-95000 Pontoise, France.
e-mail: [email protected], web: http://jecko.u-cergy.fr//
04-07-2013
Abstract
Motivated by the paper [SW] by B.T. Sutcliffe and R.G. Woolley, we present the
main ideas used by mathematicians to show the accuracy of the Born-Oppenheimer
approximation for molecules. Based on mathematical works on this approximation
for molecular bound states, in scattering theory, in resonance theory, and for short
time evolution, we give an overview of rigorous results obtained up to now. We also
point out the main difficulties mathematicians are trying to overcome and speculate
on further developments. The mathematical approach does not fit exactly to the
common use of the approximation in Physics and Chemistry. We criticize the latter
and comment on the differences, contributing in this way to the discussion on the
Born-Oppenheimer approximation initiated in [SW].
The paper neither contains mathematical statements nor proofs. Instead we try to
make accessible mathematically rigourous results on the subject to researchers in
Quantum Chemistry or Physics.

Introduction.

In the paper [SW], the authors made a remarkable effort to understand the mathematical
literature on the Born-Oppenheimer approximation. It was certainly not a easy task for
them to extract relevant information for Chemistry from papers, which often use elaborate mathematical tools and provide more or less abstract results. For instance, they
comment on the paper [KMSW], that makes use of the semiclassical pseudodifferential
calculus and of an important, but rather complicated, trick (due to Hunziker in [Hu]) to
control the Coulomb singularities appearing in the potential energy of the molecule. They
also pointed out to their colleagues in Chemistry some misunderstandings and too crude
simplications in the traditional treatment of the Born-Oppenheimer approximation for
1

Born-Oppenheimer approximation, 04-07-2013

molecules. One may feel a slightly pessimistic note in the paper [SW] on the possibility
for Chemists to use the Born-Oppenheimer approximation in a correct and accurate way
and to benefit from mathematical works on the subject. Here we shall give a description
of the situation from the mathematical point of view, which does show the present difficulties and limitations of the mathematical approach but still leads to a quite optimistic
impression. This is probably due to a different interpretation of the works [BO] and [BH].
We shall comment in details on this difference and also point out some questionable statements in the traditional presentation of the approximation.
In the last three decades, mathematically rigorous works on the validity of the BornOppenheimer approximation for molecules has been produced. We sort these works in
two (incomplete) lists into alphabetic order. The first one contains articles that, strictly
speaking, study the Born-Oppenheimer approximation: [C, CDS, CS, Ha1, Ha2, Ha3,
Ha4, Ha5, Ha7, HH, HJ1, HJ2, HJ3, HJ6, HJ8, HRJ, J1, J2, J3, JKW, KMSW, KMW1,
KMW2, Ma1, Ma2, Ma3, Ma4, MM, MS1, MS2, PST, Ra, R, ST, TW]. In the second list, we mention closely related works on semiclassical Schrodinger matrix operators:
[DFJ, FG, FR, FLN, N, Ha6, HJ4, J4, HT]. In the first list, the papers essentially show
that a reduced Hamiltonian (a Schrodinger operator with a matrix or operator valued
potential) is a good approximation to the true molecular Hamiltonian. In the second list,
the works obtain mathematical results on the reduced Hamiltonian, that are of physical
or chemical relevance for molecules.
In the present paper, we focus on the Born-Oppenheimer approximation in the mathematical sense, namely the possibility to approximate, for large nuclear masses, the true
molecular Hamiltonian by some effective Hamiltonian usually called the adiabatic operator. In Section 2, we review the original Born-Oppenheimer approximation and its
actual version in Physics and Chemistry. In Section 3, we proceed to the removal of the
centre of mass motion in two ways, one adapted to the study of bound states and the
other to scattering theory. In Section 4, we present the core of the mathematical form
of the Born-Oppenheimer approximation and describe the construction of the adiabatic
Hamiltonian. In Section 5, we compare this approach with the usual interpretation and
extract the main differences. In Section 6, we explain selected mathematical works and
comment on the actual difficulties and limitations of the theory. In Section 7, we sum up
the main features in the mathematical Born-Oppenheimer approximation and argue that
further progress towards chemically relevant questions can be reasonably achieved. The
last section is followed by an appendix that provides explanations on some mathematical
notions, that might be not familiar to the reader. Finally, we added two figures used at
many places in the text.
As pointed out in the abstract, we only present intuitive arguments and statements, that
do not respect at all the standard rigour in mathematics. But they do have a rigorous
counterpart in the mathematical literature.
Acknowledgement: The author is particularly grateful to S. Golenia for the figures. He
also thanks V. Georgescu, B.T. Sutcliffe, and R.G. Woolley for fruitful discussions. The
author is financially supported by the CNRS and the ANR NOSEVOL.

Born-Oppenheimer approximation, 04-07-2013

Review of the Born-Oppenheimer approximation

in Physics and Chemistry.

In this section we shortly review the framework and the meaning of the Born-Oppenheimer
approximation, as it is presented in Physics and Chemistry. Our review is based on
the references [BH, Me, SW, We]. We also comment on and criticize this presentation,
preparing in this way the comparison with the mathematical approach (developed in
Section 4) we shall perform in Section 5. For simplicity, we do not remove the motion
of the centre of mass. This can however be done: see [SW]. To begin with we recall the
original work [BO]. Then we focus on the actual theory.
We consider a molecule with M nuclei with positive masses m1 , m2 , , mM respectively,
with positive charges Z1 , Z2 , , ZM respectively, and with N electrons with mass set
equal to 1. We set the Planck constant ~ and the electronic charge e to 1. We denote
by (xn )1 , , (xn )M the positions of the nuclei and the corresponding momentum operators by (Pn )1 , , (Pn )M . We set xn = ((xn )1 , , (xn )M ) and Pn = ((Pn )1 , , (Pn )M ).
Similarly the positions and momentum operators of the electrons are denoted by xe =
((xe )1 , , (xe )N and Pe = ((Pe )1 , , (Pe )N ), respectively. The Hamiltonian of the
molecule is given by:
Hmol = Tn + Te + W ,
N
M
X
X
1
1
2
(Pn )i , Te =
(Pe )2i ,
Tn =
2mi
2
i=1
i=1
X
X
Zi Zj
+
W =
|(xn )i (xn )j |

i,j{1, ,N }
i6=j

i,j{1, ,M }
i6=j

M X
N
X
i=1 j=1

(2.1)
(2.2)
1
|(xe )i (xe )j |

Zi
.
|(xn )i (xe )j |

(2.3)

We consider the eigenvalue problem Hmol = E. In [BO] the authors introduce a small
parameter which is an electron/nucleon mass ratio to the power 1/4 and consider the
clamped nuclei Hamiltonian Te + W , where xn is viewed as a parameter. They assume
that, for all xn , there is a simple eigenvalue (xn ) of the clamped nuclei Hamiltonian
for the nuclear position xn . Let xe 7 (xn ; xe ) be an associated eigenvector. The set
{(xn ); xn R3M } is a potential energy surface (PES). They make the assumption of
classical nature that the nuclei stay close to some isolated equilibrium position. Such a
position x0n turns out to be an isolated, local mimimum of the PES. Notice that (x0n , 0) is
also an isolated, local mimimum for the nuclear, classical Hamilton function
M
X
1
(xn , pn ) 7 h(xn , pn ) =
(pn )2i + (xn ) .
2mi
i=1

(2.4)

Therefore an isolated equilibrium position can be identified with an equilibrium position

in the sense of classical mechanics. Near such an equilibrium position, the authors perform an expansion of Taylor type w.r.t. of the quantum eigenvalue equation. Assuming

Born-Oppenheimer approximation, 04-07-2013

the existence of an expansion in power of of the eigenvalue and the corresponding eigenvector, they compute many terms of them. The eigenvalue E is then close to the value of
at the equilibrium position.
Now we recall the modern version of the Born-Oppenheimer approximation in Physics
and Chemistry. We mainly follow the book [BH] and add some contributions from
[Me, SW, We]. The starting point is again the construction of the electronic levels (or
PES). One considers the clamped nuclei Hamiltonian Te + W and regard the nuclear variables xn as parameters. Fix such a position xn . Let (a (xn ))a be the eigenvalues of
Te + W (xn ), repeated according to their multiplicity. It is assumed that one can solve the
eigenvalue problem in the sense that, for all eigenvalue a (xn ), we have an eigenvector
xe 7 a (xn ; xe ). This is done for all possible xn . Based on the commutation of clamped
nuclei Hamiltonian with nuclear position operators and on a simultaneous diagonalization of these operators, it is then argued that the family (a (; ))a is complete and that
the bound state can be recovered by a superposition of electronic levels:
X
(xn ; xe ) =
a (xn )a (xn ; xe ) ,
(2.5)
a

It is also argued that, for each xn , the family (a (xn ; ))a can be chosen orthonormal for
the scalar product in the xe variables. Inserting (2.5) in (Hmol E) = 0 and taking the
scalar product in the xe variables with b , one derives the following coupled equations
indexed by b:
X
(Tn b + (b E)b )b +
ca,b a = 0 ,
(2.6)
a

where the terms ca,b involve derivatives of the a w.r.t. xn and the vector of nuclear
momentum operators Pn .
Making the Born-Oppenheimer approximation in Physics corresponds to the assumption
that one may neglect the last term on the l.h.s of (2.6). The equations become uncoupled
and a bound state may be written as a product of the form b (xn )b (xn ; xe ). A justification
of such an approximation is provided in [Me] when the electronic levels are simple and
sufficiently separated. One can also assume that a part of the contribution of the ca,b is
negligible (cf. the first order Born-Oppenheimer correction).
In [We], Equation (2.6) is derived in a slightly different way using states with definite
values of the electronic and nuclear coordinates. An interesting point is a discussion of
the original result by [BO] in the framework of Equations (2.5) and (2.6).

In [SW], the authors expressed their dissatisfaction in the above, traditional derivation
of the Born-Oppenheimer approximation and claimed that (2.6) is not properly justified.
They did not capture the main difficulties but felt one of them, namely the absence of
normalized eigenvectors for continuous eigenvalues.
If a (xn ) is a true eigenvalue, the corresponding eigenvector xe 7 a (xn ; xe ) is a square
integrable function in the xe variables that satisfies
(Te + W (xn ))a (xn ; ) = a (xn )a (xn ; ) .

(2.7)

But the spectrum of Te + W (xn ) contains a continuous part and, if a (xn ) sits in this part,
then Equation (2.7) has (most of the time) no nonzero square integrable solution. It is not

Born-Oppenheimer approximation, 04-07-2013

explained, neither in [BH] nor in [Me, SW, We], in which sense a (xn ; ) is an eigenvector
when a (xn ) belongs to the continuous spectrum. In [We] one uses joint eigenvectors
for the electronic and nuclear positions operators, which have only continuous spectrum.
There is an intuitive description of these states (see Section 3.2 in [We]) and their properties are somehow postulated, inspired by the true properties that one has when the
operator has discrete spectrum (like the harmonic oscillator). The main drawback with
this notion of eigenvectors is its vagueness. One cannot check their properties, one cannot find out in which meaning any formula involving them has to be understood. For
instance, with a ranging in a set having a continuous part, what is the sum in (2.5)? Even
worse, the reader is not warned that there is some fuzziness at that point. But it would
be easy to write something like: Imagine that the element in the continuous spectrum
have eigenvectors with the following properties ... and then derive Equations (2.5)
and (2.6). Or, even better, one could replace the electronic Hamiltonian by an harmonic
oscillator and perform the derivation in that case. This would at least give an idea of the
Born-Oppenheimer approximation without introducing confusion.
It turns out that many notions and statements in the above description can be reformulated in a clear, coherent, and rigorous way, and the main tool for that is the complicated,
subtle notion of spectral resolution (see the Appendix). In particular, the latter encodes
in a correct way the influence of the continuous spectrum. This will be seen in Section 4
and commented in Section 5.
Nevertheless there is another, much deaper difficulty that might hinder the derivation of
(a precise version of) (2.6). It is the presence of thresholds in the spectrum of Te +W (xn )
between the bottom of its continuous part and the energy 0. In Section 5, we shall try to
explain the notion of threshold and its influence on (2.6).
Our latter comments seem to reinforce the pessimistic posture in [SW]. But we shall see
that, thanks to the mathematical approach presented in Section 4, one can avoid complicated issues concerning the continuous spectrum and the spectral resolution and present
the most useful part of the Born-Oppenheimer approximation in a rather elementary way.

Removal of the mass centre.

In this section, we prepare the presentation of the mathematical version of the BornOppenheimer in Section 4. We recall the framework and perform the removal of the
centre of mass.
For convenience, we change the notation of the variables and rewrite (2.1), (2.2), and (2.3).
Denoting by z1 , z2 , , zM R3 the positions of the nuclei and by zM +1 , zM +2 , , zM +N

Born-Oppenheimer approximation, 04-07-2013

R3 the positions of the electrons, the Hamiltonian of the molecule is given by:
Hmol = K + W ,
M +N
M
X
1 X
1
zj ,
zi
K =
2m
2
i
i=1
j=M +1
X
X
Zi Zj
+
W =
|zi zj |

(3.2)

i,j{M +1, ,M +N }
i6=j

i,j{1, ,M }
i6=j

(3.1)

1
|zi zj |

+N
M M
X
X
i=1

Zi
.
|zi zj |
j=M +1

(3.3)

Here zk denotes the Laplace operator in the zk = (zk1 , zk2 , zk3 ) variable, that is
zk = 2zk =

izk1

2

+ izk2

2

+ izk3

2

where t stands for the partial derivative with respect to the variable t.
It is usual and physically relevant to remove from the Hamiltonian Hmol the motion of
the centre of mass of the molecule. This is done by an appropriate change of variables.
There is no canonical choice for this change of variables. This means in particular that
one can choose it according to the kind of study one wants to perform. To study bound
states of the molecule or its time evolution, we shall use the change of variables adopted
in [KMSW, MS2, SW]. To consider diatomic collisions (ion-ion, ion-atom, or atom-atom
scattering), we shall use another one (those in [KMW1, KMW2]); see p. 75-82 in [RS3]
for details on the removal of the centre of mass.
In the first mentioned situation, we take the nuclear centre of mass (which is close to the
centre of mass of the molecule), Jacobi coordinates for the nuclei, and atomic coordinates
for the electrons. Let C : (z1 , , zM +N ) 7 (R; x1 , , xM 1 , y1 , , yN ) be the change
of variables defined by
M
X

1 X
mk zk , for 1 j N , yj = zM +j R ,
m
k=1
k=1
X
1
for 1 j M 1 , xj = zj+1 P
mk zk .
kj mk kj

m =

mk , R =

R is the centre of mass of the nuclei, the xj are the new nuclear coordinates, and the
yj are the new electronic variables. For an appropriate constant C that only depends on
the masses and on N we define, for any L2 function f of the variables (z1 , , zM +N ),


(Uf )(R; x1 , , xM 1 , y1, , yN ) = Cf C 1 (R; x1 , , xM 1 , y1 , , yN ) . (3.4)
The constant C is chosen such that, for all f , f and Uf have the same L2 -norm (U
is unitary), keeping unchanged the physical interpretation of the L2 -norm. Looking at

Born-Oppenheimer approximation, 04-07-2013

the Hamiltonian Hmol in the new variables (R; x1 , xM 1 , y1 , , yN ) means that we

consider the operator
1
UHmol U 1 =
R + H ,
2m
where H only acts on the variables (x1 , , xM 1 , y1 , , yN ). Forgetting about the
kinetic operator of the nuclear centre of mass, we focus on the physically relevant Hamiltonian H. Denoting by t the gradient operator in the variable t and setting,
X 1
1
1
mk
, x = (x1 , , xM 1 ), and y = (y1 , , yN ),
for 1 j M 1, j = mj+1 +
kj

the latter is given by

H = H0 + THE ,
M
1
X
1
xj + Q ,
H0 =
2
j
j=1
X
THE =
ckj yk yj ,

(3.5)
(3.6)
(3.7)

1k<jN

Q =

Q(x) dx , Q(x) =

N
X

yk + W (x; y)

(3.8)

k=1

where W (x; y) is just the function W in (3.3) composed with the inverse change of variables
C 1 . The action of the direct integral operator Q on a function (x; y) gives another
function (x; y) defined, for all x, by (x; y) = Q(x)(x; y), that is the action of the
operator Q(x) on the function y 7 (x; y). Details on the notion of direct integral are
provided in the Appendix.
Compared to Section 2, the x (resp. y) variables play the role of the xn (resp. xe )
variables. We observe that Q(x) is an operator in the electronic y variables that depends
only parametrically on the nuclear x variables and does not depend on R. The operator
THE is usually called the Hughes-Eckart term. For each nuclear configuration x, the
operator Q(x) is refered to as the electronic Hamiltonian in the configuration x (it is
called the clamped-nuclei Hamiltonian in [SW]). Note that the coefficients j in H are
missing in [KMSW]. This has no consequence on the validity of the results in this paper,
that also hold true for the present Hamiltonian H.
Next we turn to the scattering situation. For simplicity, we restrict ourselves to the
diatomic case (i.e. M = 2). We still look at the Hamiltonian Hmol but we want now to
describe the collision of two ions (or two atoms, or an atom and an ion). It is useful to
choose a change of variables, that allows a easy description of the system at the beginning
of the collision process (and another one, to describe the system after the collision). To
this end, we introduce a cluster decomposition c = {c1 , c2 } with cj = {j} cj , for j = 1,2,
and c1 , c2 form a partition of the set {3, , N + 2}. At the beginning of the scattering
process, the particles are gathered in two clusters described by c1 and c2 . Each cluster
constains a nucleus illustrating the fact that we consider a collision of two ions (and not
a collision of some electrons with a molecule). Since the motion of the centre of the

Born-Oppenheimer approximation, 04-07-2013

system is not relevant for scattering, we shall remove it. In order to do so, we use the
particular change of variables in [KMW1, KMW2], which also allows a good description
of the scattering processes associated to the decomposition c.
For k {1; 2}, denote by |ck | the number of electron in the cluster k. Its mass is then
mk := mk + |ck | and its mass centre is located at:
X 
1 
zj .
(3.9)
Rk :=
mk zk +
Mk
jc
k

In particular, the total mass m of the molecule is m = m1 + m2 + N = m1 + m2 . The

new variables are, for k {1; 2},
N +2

X 
1
m1 z1 + m2 z2 +
zj , x := R1 R2 , yj := zj zk , for j ck .
R :=
M
j=3
We set, for y R3N ,
(y) :=

1 X
1 X
yj
yj .
M1 jc
M2 jc
1

(3.10)

(3.11)

As above, this change of variables Cc induces a unitary map Uc by

(Uc f )(R; x, y1, , yN ) = Cc f Cc1 (R; x, y1 , , yN ) .

(3.12)

Looking at Hmol in the new variables amounts to considering

Uc Hmol Uc1 =

1
R + H ,
2m

where H only acts on the variables (x, y1 , , yN ). Taking away the motion of the centre
of mass of the full system again, we keep our attention on H , which is given by
1 1
1 

x + THE
+ Qc ,
(3.13)
H :=
+
2 m1 m2
2
2
X
1 X


i
(3.14)
THE
=

yj 
2m
k jc
k=1
k
Z
Z
Qc =
Qc (x) dx , Ic =
Ic (x) dx ,
(3.15)
Qc (x) := Qc + Ic (x) ,
(
)
2

X 1
X
X
1
Z
k
yj
Qc :=
+
,
2
|yj |
|yi yj |
jck
k=1
i,jck ,i6=j
X
X
Z1 Z2
Z1
Z2

+
Ic (x) :=
|yj + x (y)| jc |yj x + (y)| |x (y)|
jc1
2
X
X
1
1
+
,
+
|y

y
+
x

(y)|
|y

x
+
(y)|
i
j
i
j
ic , jc
ic , jc
1

Hc

:= H Ic .

(3.16)

(3.17)
(3.18)

(3.19)

Born-Oppenheimer approximation, 04-07-2013

Qc stands for the Hamiltonian of separated (noninteracting) clusters while Ic contains all
extracluster interactions. The electronic Hamiltonian for the nuclear position x is Qc (x).

The term THE

is the Hughes-Eckart term in this situation. As usual in scattering theory,
we should mention the reference (or free) dynamics to which the full dynamics has to be
compared (for time tending to ). The free dynamics is generated by the Hamiltonian

Hc in (3.19), that is, up to the Hughes-Eckart term THE

, the Hamiltonian of freely moving
clusters c1 and c2 . If the scattering process under consideration produces at the end (i.e.
for time tending to +) a cluster decomposition d of the system then the free dynamics
for large positive times is given by Hd .
Using Katos pertubation argument and Hardys inequality (cf. [K, RS2]), one can show
that H can be realized as a self-adjoint operator. This means that H can be defined on an
appropriate subspace of the L2 functions of the variables (x1 , , xM 1 , y1 , , yN ) and
the resulting operator is self-adjoint. In fact, this subspace is the domain of the Laplace
operator in all variables (x1 , , xM 1 , y1, , yN ). The same holds true for H0 . In the
same way, Hmol (respectively H ) is self-adjoint on the domain of the Laplace operator
of all variables z1 , , zM +N (respectively x, y1 , , yN ). For fixed x, Q(x) (respectively
Qc (x)), as an operator on the variables y1 , , yN , also has a self-adjoint realization by the
same argument. Since Q (respectively Qc ), as an operator on the variables x, y1 , , yN ,
is a direct integral of self-adjoint operators, it has also a self-adjoint realization (cf. p.
284 in [RS4]).

Mathematical approach to the Born-Oppenheimer

approximation.

Here we want to present the main ideas behind the mathematical treatment of the BornOppenheimer approximation, which was initiated in [C, CDS, CS]. Since its validity
should rest on the fact that the nuclei are much heavier than the electrons, one introduces
a small, positive parameter related to the electron/nucleon mass ratios. For instance, in
[KMSW], the nuclear masses mk are given by mk = 2 k , where the k are of order 1, and,
in [KMW1, KMW2], one uses 2 = M11 + M21 (with the notation of Section 3). Anyhow
the main point is that is always sent to 0. This means that the results proved hold true
for small enough and, most of the time, one has no concrete idea of how small should
be. This restriction is of course a drawback for physical or chemical purposes but it is
useful to understand the small limit and it often gives correct results when compared
with the observed behaviour of the physical system.
Now we come to the main features that ensure the validity of the Born-Oppenheimer
approximation. Let us consider a normalized, bound state of energy E of the operator
2 1
H in (3.5). We note that 1
for -independent j . Let us temporarily fix the
j = j
nuclear variable x. Typically the spectrum of Q(x) starts with some isolated eigenvalues
1 (x), , J (x) with finite multiplicity k1 , , kJ respectively and has above a continuous part c (Q(x)) (see fig. 1 in the diatomic case). Here J could be infinite, i.e. one may
have an infinite number of eigenvalues below c (Q(x)). Since Q(x) is self-adjoint, we can

10

Born-Oppenheimer approximation, 04-07-2013

decompose (x; ) in a basis of electronic eigenvectors of Q(x):
(x, ) =

kj
J X
X

hjk (x; ), (x; )iy jk (x; ) +

dSx ()(x; ) ,

(4.1)

inf c (Q(x))

j=1 k=1

where the jk (x; ) form a basis of true eigenvectors of Q(x) associated to j (x) respectively, h, iy denotes the usual scalar product in the y variables, and dSx () is the (restriction to the continuous spectrum of the) spectral resolution of Q(x) (see the Appendix
for details). This can be done for all values of the variable x. We assume that E belongs
to a small open energy interval (E ; E+ ) below the infimum over all x of c (Q(x)) (as
in fig. 1). Since the coefficients ckj in (3.7) contain 2 , the Hughes-Eckart term THE is
small compared to the electronic Hamiltonian Q in (3.8). To compute E and , the idea
is that only the part of the spectrum of the operators Q(x) less or equal to E+ should be
relevant, since the nuclear kinetic energy is nonnegative. Define J+ as the largest j J
such that there exists some x with j (x) E+ . In (4.1), we expect that
(x, ) =

J + kj
X
X

hjk (x; ), (x; )iy jk (x; ) + small term .

(4.2)

j=1 k=1

Let (x) be the orthogonal projection on the first J+ energy levels of Q(x) that is, for an
electronic wavefunction (y),
(x) =

J + kj
X
X

hjk (x; ), iy jk (x; ) .

(4.3)

j=1 k=1

We set (x) = 1 (x). Note that (4.1) may be rewritten as

(x, ) = (x)(x, ) + (x)(x, ) ,

(4.4)

where

(x)(x, ) =

Z
kj
J
X
X
hjk (x; ), (x; )iy jk (x; ) +

j=J+ +1 k=1

dSx ()(x; ) .
inf c (Q(x))

To implement this idea, it is natural to try as an effective, self-adjoint Hamiltonian an

operator of the form G where G may be H or H0 and is the direct integral of the
(x). It acts on a total wavefunction (x, y) as follows: for each x, one projects (x, )
as in (4.3) then one lets G act on the result and, finally, one projects again according to
(4.3). Thus, the spectral subspaces of the operators Q(x) corresponding to energies above
E+ are removed. Such an operator is usually called an adiabatic operator. We expect
that among the eigenvalues in (E ; E+ ) and corresponding eigenvectors of G, there is
a good approximation of E and . Let us try to justify intuitively this claim.
Assume that (E ; E+ ) is a small interval around the infimum of the function 1 that is
attained in some region (in the diatomic case, 1 only depends on the norm of x and
can be a sphere) and that J+ = 1 (for instance (E ; E+ ) = (E0 ; E+0 ) in fig. 1). Using

Born-Oppenheimer approximation, 04-07-2013

11

the positivity of the nuclear kinetic energy, H Q + THE thus, by the properties of direct
integrals (see the Appendix) and the identity (x) + (x) = 1,
Z
E = h, Hi
h(x, ), Q(x)(x)(x, )iy dx
Z
+ h(x, ), Q(x) (x)(x, )iy dx
Z
+ h(x, ), THE (x, )iy dx .
Since THE is small compared to Q, one can show via Katos pertubation theory that
Z


h(x, ), THE (x, )iy dx = O(2 ) and THE = O(2 ) ,
(4.5)

where k k is the L2 -norm in the variables (x, y1 , , yN ). Therefore, since 1 (x) is simple,
Z

2
E
1 (x)h(x, ), 1(x; )iy  dx
Z
+ E+ k (x)(x, )k2y dx + O(2 ) ,
(4.6)
where k ky is the L2 -norm in the y variables. Here we used the fact that, on the range of
(x), Q(x) E+ . Since E E+ and, by (4.4) and the properties of direct integrals,
Z
Z
2

2


k (x)(x, )k2y dx ,
(4.7)
1 = kk =
h1 (x; ), (x, )iy dx +
2

the second integral in (4.6) and the term h1 (x; ), (x, )iy  for x far from should be
small. In particular, in (4.1), we should have
(x, ) = (x)h1 (x; ), (x; )iy 1 (x; ) + small term ,

where is the characteristic function of a neighbourhood of . The relevant part of

is then its projection onto the electronic level 1 (x; ). Since the lower bound in (4.6) is
close to E and does not contain the nuclear kinetic energy, the latter must be small.
Note that this is not sufficient if the interval (E ; E+ ) is placed much higher (as in fig. 1).
Let us now produce a better argument in this more general situation. Using the second
estimate in (4.5), one can show that, close to E and respectively, there are an energy
E0 and a normalized L2 -function 0 such that H0 0 = E0 0 . We assume E0 (E ; E+ ).
In particular, (H0 E0 )0 = 0 and (H0 E0 )0 = 0. Using + = 1 and 2 = ,
(H0 E0 )0 = H0 0 ,
( H0 E0 ) 0 = H0 0 .
Since (x) (x) = 0 and (x) commutes with Q(x),

H0

1
M
1
M

X
X
1
1

2
=
[xj , ] .
xj =

2
2
j
j
j=1
j=1

(4.8)
(4.9)

Born-Oppenheimer approximation, 04-07-2013

12

Now 2 times the commutator [xj , (x)] equals 2(xj )(x) xj + 2 (xj )(x). Since
Z

2
1 
xj 0  dx
2j

Z M
1
X
2
1 
xk 0  dx
2k
k=1

and the nuclear kinetic energy remains bounded (this is due to the self-adjointness of H0 on
the domain of the Laplace operator and to the finitness of the total energy), the right hand
sides of (4.8) and (4.9) are O() in L2 -norm. Since H0 E+ on the range of ,
the operator H0 E0 restricted to this range is invertible with bounded inverse and
(4.9) shows that 0 is O() in L2 -norm. In particular, 0 is almost an eigenfunction
of H0 with L2 -norm close to 1 (cf. (4.7)) so is close to a true normalized eigenfunction
of H0 . Thus 0 (and also ) should be also close to a normalized eigenfunction of
H0 . Notice that, by (4.8), the nuclear kinetic energy in the state 0 is close to
Z
h0 , (Q(x) E0 )0 iy dx
which has no reason to be small in general. Indeed, if E0 is clearly above the infimum of
1 and below the infimum of 2 , J+ = 1 and this term equals
Z

2
(1 (x) E0 )h1 (x; ), 0 (x, )iy  dx

and it is possible to show that the squared function essentially lives in the bottom of the
well: {x; 1 (x) E0 }.
In the above argument, we used the fact that we can differentiate (x) twice with respect
to the variable x. This is not obvious at all when one looks at the x-dependence in
Q(x) (see (3.8)) which involves the rather irregular function W in (3.3). Thanks to a
trick due to Hunziker in [Hu], one can prove that x 7 (x) is smooth away from the
set of the nuclear collisions (this is actually sufficient for our argument above since one
can show, by energy arguments, using the repulsive nature of nuclear interaction, that
the wavefunction is concentrated away from these collisions). This trick is also used in
[KMSW] and is partially responsible for the technical complications there. The idea is
to perform a x-dependent change of variables on the y variables in Q(x) that makes the
x-dependent singularities in the function W (x, y) in (3.8) x-independent. This does not
change the (x). This can be done only locally in x i.e. for x close enough to any fixed
position x0 (see Lemma 2.1 in [J5] for details). It is essential that the j (x) are separated
from the rest of the spectrum of Q(x) (in particular from its continuous part). This
implies the finitness of the rank of (x) but this property is not used.
The regularity of (x) allows the contruction of globally defined, smooth functions x 7
jk (x, ) (for 1 j J+ and 1 k kj ) with values in the L2 electronic functions such
that, for each x, the family (jk (x, ))jk is a orthogonal basis of the range of the projection
(x). Because of the possible presence of eigenvalue crossing (see fig. 1), it is not always
possible to choose these functions jk (x, ) among the eigenvectors of Q(x).
We have seen that, up to an error of size O(), one can reduce the eigenvalue problem
for H to the one for H0 . One can compute explicitly H0 in terms of the electronic

Born-Oppenheimer approximation, 04-07-2013

13

wavefunctions jk (x, ) (see [PST]) and again remove terms that are also O(). Writing
as
the approximate eigenstate of energy E
=

J + kj
X
X

jk (x)jk (x; ) ,

(4.10)

j=1 k=1

one ends up, in the diatomic case for simplicity, with the
and letting act H0 E,
if the jk (x; ) are eigenuncoupled equations 2 x jk + j jk = Ejk jk , Ejk close to E,
vectors of Q(x). If the latter is not true (it can be the case when crossing eigenvalues
occur), one has coupled partial differential equations in the x variables for the jk with
coefficients depending on the hjk , Q(x)j k iy . These differential equations play the role
of (2.6) in the physical setting.
Now, if we demand an accuracy of O(2 ), then H0 still provides a good approximation
if the rank of (x) is one. Taking the electronic wave function 1 real, one cancels the
term containing (x)(xj )(x) in H0 , producing an error of order 2 . But now, less
terms in H0 can be removed. In particular, one has to keep terms containing the socalled Berry connection, i.e. factors of the form hjk , xp j k iy . We see that the variation
of the j k (or of (x)) has to be taken into account. We refer to [PST] for details.
If one wants to improve the accuracy to O(3 ) (or better), one needs to include the HughesEckart term THE in Q(x) and to replace (x) by an appropriate refined projector r (x)
which is essentially of the form (x) + 1 (x) + 2 2 (x). Then one uses as adiabatic
operator r Hr (cf. [MS2, PST]).
In the scattering situation mentioned in Section 3, one can choose the total energy E
in an energy range (E+c , Ec ) like in fig. 2 and take H as an effective Hamiltonian.
In this case, it is important to choose the projections (x) as spectral projections of the

operators Qc (x)+THE
(because the Hughes-Eckart term THE
has no decay in x at infinity).

Comparison.

Having presented the physical and mathematical versions of the Born-Oppenheimer approximation in Sections 2 and 4 respectively, we are now ready to compare them. After a
remark on a minor difference, we shall explain how to partially cure the vagueness of the
traditional approach and then focus on the comparison.
While, in Physics, the small parameter is essentially an electron/nucleon mass ratio to
the power 1/4, we chose here more or less the square root of this ratio. This comes from
the fact that mathematicians see the Born-Oppenheimer approximation as a semiclassical
analysis w.r.t. the nuclear variables and, in this framework, it is convenient to consider
-derivatives: x in the nuclear variables. Since we fixed the total energy, we consider, roughly speaking, 2 x of order one. So is also x . In particular, a term like
2(xj )(x)xj is of order (and this can be shown rigorously, in an appropriate sense).
This gives a way to detect negligible terms. Of course, the choice of is a matter of taste,
since we could have replaced everywhere by 2 .

14

Born-Oppenheimer approximation, 04-07-2013

Notice first that the presentations in Sections 2 and 4 are slightly different since we removed the centre of mass motion in the latter and not in the former. However this
will not influence the discussion below. For their comparison, we shall use the framework of Section 4 making the following identifications: xn x, xe y, Te + W Q,
Te + W (xn ) Q(x), Tn H0 Q.
In [BH, We], a complete, orthogonal set of eigenfunctions is considered and used to
express the total wave function as a superposition of such eigenfunctions. In this set,
there are true eigenfunctions and their orthogonality is expressed in term of the L2 scalar
product. As in the well-known case of the hydrogen atom, the set of true eigenfunctions
is not complete. One has to add a contribution of the continuous spectrum. But it is not
clear in [BH, We] what are the (generalized) eigenfunctions associated to energies in the
continuous spectrum. The properties of this set (completeness and orthogonality) are deduced from the self-adjointness of the electronic Hamiltonian, viewed in the full L2 -space,
and its commutation with the nuclear position operators. Note that the commutation of
unbounded self-adjoint operators is a subtle issue (see [RS1], p. 270-276). Fortunately
we shall not encounter a difficulty of this kind here. But the definition of (generalized)
eigenfunctions has to be clarified. In fact, the rather involved notion of spectral resolution
associated to the electronic Hamiltonian (see the Appendix) is hidden behind this.
Let us start with a simple situation. We consider the multiplication operator T by the
real variable t in L2 (Rt ) (this is a toy model for the position operators studied in [We]).
One can verify that T is self-adjoint and compute its spectral resolution S. The latter is
the identity operator on L2 (Rt ) times the Lebesque measure dt on R (that is the measure
that associates to any finite interval its length). If is such an interval and L2 (Rt )
then S() = || , where || is the length of . The completeness of S is expressed by
the formula
Z
=
dS() ,
R

which corresponds to (2.5) and to (4.1). The vectors in the range of S() for all possible
play the role of the electronic eigenvectors. There is a kind of orthogonality. If and
are disjoint subsets in R then, for any , L2 (Rt ), it turns out that
hS() , S( ) i = 0 .

(5.1)

For any fixed R, S({}) = 0 since the length of the interval {} is zero. This means
that is not an eigenvalue of T . In this framework, there is no state with exact position
and can be interpreted by saying that the classical notion of exact position is not wellsuited in a quantum situation. One can try to seek an eigenvector outside the Hilbert
space L2 (Rt ). One finds a Dirac distribution at , denoted by (t ). This is not a
function! The product of two distributions is not defined in general. In particular, the
square of a Dirac distribution and therefore
Z
h(t ) , (t )i =
(t )2 dt
R

are not defined. So what is the meaning of Equations (3.2.14) and (5.6.8) in [We]? Thus
the introduction of states with definite electronic and nuclear positions in [We] does require a more precise definition.

Born-Oppenheimer approximation, 04-07-2013

15

Let us come back to the physical version of the Born-Oppenheimer approximation. For
each nuclear position x, Q(x) is a self-adjoint operator. Therefore it has a spectral resolution Sx supported in its spectrum (Q(x)) (see the Appendix). The latter can be split
into two parts: a discrete one and above a continuous one. The discrete part consists
of eigenvalues of finite multiplicity. The continuous part c (Q(x)) is an interval [tx ; +[
where tx := inf c (Q(x)) 0. The operator Q(x) have no positive eigenvalues but may
have eigenvalues in [tx ; 0] (see [CFKS]). The latter are called embedded eigenvalues (in
the continuous spectrum). The closed interval [tx ; 0] also contains so-called thresholds.
At such a point the nature of the spectrum changes. For instance, tx is a threshold which
separates discrete and continuous spectra. 0 is also a threshold at which a change of
multiplicity of the spectrum occurs. One can have such thresholds between tx and 0.
The number of eigenvalues below tx may be infinite. This is known as the Efimov effect
(see [Wa2, Y]). In this case, the eigenvalues accumulate at tx . The (possible) embedded
eigenvalues have finite multiplicity and can accumulate only at thresholds (cf. [ABG]).
This complicated structure is reflected in the spectral measure. The discrete eigenvalues
below tx produces the first term on the r.h.s of (4.1), that is a sum of the orthogonal projections in the y variables hjk (x; ), )ijk (x; ). From the second term, one can extract
the contribution of the embedded eigenvalues producing in this way similar terms to the
first one. In the rest, which is an integral over [tx ; +[, one can write, outside the set of
thresholds, dSx () = Sxd ()d for some operator-valued function 7 Sxd () (see [ABG]).
The superscript d refers to density since Sxd is, away from the thresholds, the density of
the measure dSx w.r.t. the Lebesgue measure d. It can be interpreted as the derivative
w.r.t. of 7 Sx (] ; ]). Denote by Tx the set of thresholds of Q(x). One can
split the set of embedded eigenvalues of Q(x) in a disjoint union of sets Ax (t) for t Tx ,
such that the elements of each Ax (t) can only accumulate at t. For Ax (t) denote
by k() its multiplicity and by ;1 (x; ), , ;k() (x; ) a set of normalized eigenvectors
associated to . By the above computation we may rewrite (4.1) as
(x, ) =

kj
J X
X

hjk (x; ), (x; )iy jk (x; )

j=1 k=1

k()
X X X

h;k (x; ), (x; )iy ;k (x; )

(5.2)

tTx Ax (t) k=1

tx

Sxd ()(x; ) d .

Note that in the integral in (5.2), the integrand is defined outside the discrete set Tx .
(5.2), as well as (4.1), is a correct superposition of electronic states (for a fixed nuclear
position at x).
In the special case where Q(x) is a one-electron Hamiltonian, Tx = {0} and one expects
to be able to construct generalized eigenfunctions in the continuum in order to express
the operators Sxd (). This has been done in a similar situation (see [RS3], p. 96). Thus
we would get a justified version of (2.5). Note that the indices a in the latter such that
a = 0 would be absent, if 0 is not an eigenvalue, and would contribute to the second
term on the r.h.s. of (5.2), if 0 is an eigenvalue. Even in this simple case, we do not know

Born-Oppenheimer approximation, 04-07-2013

16

the behaviour of Sxd near 0. There exist some results on this question (see [JK]) but not
exactly for the electronic Hamiltonian we consider here. For a many-body Hamiltonian
Q(x) we a priori have the full complexity of (5.2) and again the behaviour of Sxd near Tx
is not known. See however [Wa1] in a similar situation. Furthermore we are not aware of
a construction of generalized eigenfunctions in the continuum for N-body systems.
To get a rigorous version of (2.6), we could try to let H E act on (5.2). To this end we
would essentially need to differentiate w.r.t. x many terms in (5.2) like tx , Sxd () but also
the sets Tx and Ax (t). This would probably be a quite involved task, if it is at all possible
in some sense. Assume it is. Then we can take the scalar product in the y variables with
jk (x; ) and probably get an equation for jk (x) = hjk (x; ), (x; )iy similar to (2.6),
using the fact that jk (x; ) is orthogonal to the last two terms in (5.2). This should work
also if we take the scalar product with ;k (x; ) for the same reason. In both cases, the
terms corresponding to the ca,b in (2.6) would be more complicated. Now, to complete
the analogy with (2.6), we would seek an equation for Sxd ()(x; ) or perhaps Sxd (), for a
fixed value tx . To our best knowledge, it is however not known if the range of Sxd () is
orthogonal to the one of Sxd ( ) if and are different. We do have here a weak form of
orthogonality similar to (5.1) but it is not clear that it is sufficient to extract an equation
for Sxd ()(x; ) or Sxd () for fixed . In the above strategy, we neglected the difficulty of
the control of Sxd () and its x-derivatives up to second order near the thresholds. But the
behaviour of Sxd at thresholds is already an open problem.
To summarize our discussion on Equations (2.5) and (2.6), we have seen that it is possible
to produce a rigorous superposition of electronic states at fixed nuclear positions, namely
(4.1) or (5.2). If one prefers a version in the full L2 space, one can use the formula
Z
=
(x; ) dx ,
(5.3)

where each (x; ) is given by (4.1) or (5.2) (cf. the Appendix). Equation (5.3) is a
well-defined, correct superposition of electronic states that can replace (2.5). Actually
(5.3) may be interpreted as the representation of in a basis of common eigenvectors
of Q and of the nuclear position operators. Along this lines, the derivation of a precise
version of (2.6) seems to be quite difficult, much more difficult than the usual, simple (formal) derivation of (2.6). Thus, if there is another, more clever construction of electronic
eigenvectors for the continuous spectrum that leads to a well-defined version of (2.5)
and (2.6), it deserves to be explained in details.
If one gives up the exact computation of the wave function , one still have the opportunity
to use the previous ideas to look for a good approximation of it. That is precisely what
we did in Section 4 with the help of an energy argument. The latter allowed us to replace
(2.5) by the approximation (4.2) (see also (4.10)), that only contains well-defined, quite
elementary terms. The controversial Equation (2.6) is then replaced by easily derived
coupled differential equations on the functions jk (see just after (4.10)). In particular, we
were able to forget about the contribution of the continuous spectrum of the electronic
Hamiltonians Q(x).
In the framework of Section 4, making the Born-Oppenheimer approximation in the
sense used in Physics and Chemistry amounts to removing off-diagonal terms from
H (or H0 ), that is to removing terms involving jk (x; ) and j k (x; ) for different

Born-Oppenheimer approximation, 04-07-2013

17

couples (j; k) and (j ; k ). In other words, the approximation is performed by deleting

in the differential equations for jk all terms involving j k for (j ; k ) 6= (j; k). This
approximation may be justified if the electronic levels j are simple and well separated
(see [Me]). This is also the case when the nuclear kinetic energy is small as in [BO],
as we saw in the first energy argument in Section 4. In general, we should expect the
breakdown of this approximation. Recall that, when the total energy belongs to (E ; E+ )
in fig. 1, we have seen that the nuclear kinetic energy is not small and we have eigenvalue
crossings, that is a change of multiplicty. In a time-dependent or scattering situation,
we even expect that the assumption of well separated electronic levels hinders significant
inelastic phenomena (see Section 6 below).
In contrast to the usual treatment of the Born-Oppenheimer approximation, we fixed in
Section 4, at the very beginning, the total energy of the system, selected certain electronic
levels (the j ) relevant for this energy, and then constructed the adiabatic Hamiltonian
that should approximate H near this energy. In the first energy argument, when the total
energy is close to the infimum of 1 , we have roughly shown that the nuclear kinetic energy
is small and recovered the result in [BO], without using the assumption that the nuclei
stay close to some position where the infimum of 1 is attained. In fact, this property
can be recovered from the analysis in [KMSW] (see comments in Section 7). It is again
an energy argument, namely the bound H0 E+ on the range of , that allows
us to show (in the limit 0), that the component of in the range of is negligible,
that is the high energy electronic levels have a small contribution to the superposition
(4.1) of electronic states that reproduces . In Section 4 we actually used (4.1) in the
form of (4.4). In particular we can recover the main issues of the Born-Oppenheimer
approximation without relying on the complicated notion of spectral resolution.
The bound H0 E+ on the range of is probably the core of the mathematical
Born-Oppenheimer approximation and plays a role in all mathematical results, that we
shall review in Section 6 below. Note that this bound crucially relies on the fact that
the nuclear kinetic energy is bounded below. If it would be described by a Hamiltonian
that is unbounded below (like a Dirac operator), we would have to expect that very high
energy electronic levels do play a role, even if the total energy of the system is fixed.

Review of mathematical results.

In this section, we present some rigorously proved results on the Born-Oppenheimer approximation that illustrate the main ideas developed in Section 4. Since we cannot review
all results, we selected one from each of the following fields: bound states, resonances,
scattering process (collision), and time evolution. These choices may be detected as abitrary (they reflect the way the author senses the subject) but we try to present results
with the highest degree of generality. Nevertheless we also comment on other results in
these fields. At the end of the present section, we add some remarks when symmetries of
the particles are taken into account.
Let us begin with the study of bound states of a molecule which was performed in the paper [KMSW] (previous results were obtained in [Ha3, Ha4]). One studies the eigenvalues

Born-Oppenheimer approximation, 04-07-2013

18

of the operator H (cf. (3.5)) in the framework introduced in Section 4. In particular, the
adiabatic operator H is used first as an effective Hamiltonian but in a slightly different way. The authors use a so-called Grushin problem and pseudodifferential techniques
to produce a more accurate effective Hamiltonian F (E) (depending on the sought after
energy E (E ; E+ )), which is a pseudodifferential matrix operator. F (E) essentially
corresponds to the operator that defines the (a priori) coupled equations on the jk we
mentioned at the end of Section 4. Then E is an eigenvalue of H (essentially) if and
only if 0 is an eigenvalue of F (E) (cf. Theorem 2.1). Here we mean that, if E is a true
eigenvalue of H, then 0 is an eigenvalue of F (E ) where |E E | = O(N ), for all integer
N, and also that, if 0 is an eigenvalue of F (E ) then H has an eigenvalue E such that
|E E | = O(N ), for all integer N. An explicit but rather complicated, infinite contruction produces the operators F (E). For practical purpose, one follows only an appropriate
finite number of steps of this construction to get an operator Fp (E) such that the above
errors are O(p ). A concrete example is given in Proposition 1.5, where eigenvalues of H
in some particular energy range are computed up to O(5/2 ).
For diatomic molecules, the authors consider an energy range close to the infimum of 1
(like (E0 ; E+0 ) in fig. 1). Recall that, for all nuclear positions x, 1 (x) is the lowest eigenvalue of the electronic Hamiltonian Q(x), which is simple. Actually, one does not need
to consider the lowest eigenvalue but it is important that it is simple and that the rank
of the projection (x) is always 1. In the mentioned energy range, the eigenvalues of H
and the corresponding eigenvectors are computed by an asymptotic expansion in power
of 1/2 of WKB type. In particular, the original result of [BO] for diatomic molecules is
contained in Theorem 3.2 in [KMSW].
For polyatomic molecules, the same situation is studied but two cases occur. Recall that
denotes the set of nuclear positions x = (x1 , , xM 1 ) R3(M 1) where the infimum of 1
is attained. It is assumed that is the set of all points (Ox01 , , Ox0M 1) where O ranges
in the set of all orthogonal linear transformations in R3 and x0 = (x01 , , x0M 1 ) .
One can check if the points x01 , , x0M 1 R3 lie on a line, or on a plane, or generates
the whole space R3 . The molecule is linear, planar, and non-planar respectively. For
a linear or planar molecule, it is shown that, in an appropriate neighbourhood of 1 s
infimum, there is exactly one eigenvalue of H which is given by a complete asymptotic expansion in 1/2 . A corresponding eigenvector can also be obtained by such an asymptotic
expansion. The distance from this eigenvalue to the rest of the spectrum of H is of order
5/2 . In the non-planar case, two different simple eigenvalues of H are present in the mentioned neighbourhood. The splitting (that is the distance between these two eigenvalues)
is exponentially small in . The eigenvalues and the corresponding eigenvectors are again
given by an asymptotic expansion in power of 1/2 . These eigenvectors can be related to
one another with the help of the reflection (x, y) 7 (x, y).
In the above framework, we mention a modification of the Born-Oppenheimer approximation performed in [HJ6, HJ7, HJ8] in order to make apparent chemical hydrogen bonds
in molecules. The main idea is to take the hydrogen mass of order 3/2 , while the mass
of the heavier atom and the electronic mass are still of order 2 and 0 = 1, respectively.
In this setting, one can reduce the eigenvalue problem to an effective one in a similar way
as in Section 4. However, the authors use a multiscale analysis as in [Ha3, Ha4].
Next we describe the paper [MM] on the resonances of the operator H in the diatomic

Born-Oppenheimer approximation, 04-07-2013

19

case. As we shall see, resonances are complex eigenvalues of an appropriate distortion of

H. It is believed that they give information for the long time evolution of the molecules
(scattering). This link has been done in other (simpler) situations but, to our best knowledge, not in the present framework, i.e. for molecules in the large nuclear masses limit.
Actually the authors consider the diatomic version of the operator H0 (cf. (3.6)), that
is the relevant molecular Hamiltonian without the Hughes-Eckart term. To explain the
announced distortion, we need some notation. Let : R3 R3 a smooth function, which
is 0 near 0 and equals the identity map (i.e. (x) = x) for |x| large. For real numbers ,
one introduces the transformation U defined on total wavefunctions (x, y) by
(U )(x, y) = |J (x, y)|1/2 (x + (x), y1 + (y1), , yN + (yN )) ,
where the function J is the Jacobian of the change of variables (x, y) 7 (x + (x), y1 +
(y1), , yN + (yN )). It turns out that one can extend U to small enough complex
values of . The distorted Hamiltonian is given by H = U H0 U1 . Since U is unitary for
real , H has the same spectrum as H0 on the real line but, for nonreal , the continuous
part of the spectrum of H is obtained from the one of H0 by some rotation in the complex
plane. Furthermore, between the continuous spectra of H0 and H , nonreal eigenvalues
of H of finite multiplicity appear. They actually do not depend on and are called the
resonances of H0 . They are close to the continuous spectrum of H0 , which is responsible
for the scattering processes governed by H0 . One wants to compute these resonances. To
this end, one faces an eigenvalue problem as above in [KMSW] but now for the non selfadjoint operator H . In [MM], it is shown that one can adapt the arguments of [KMSW]
to show that E is a resonance of a modified version of H0 if and only if 0 is an eigenvalue
of an E-dependent pseudodifferential matrix operator. Due to a technical difficulty, the
authors have to smooth out the repulsive, nuclear interaction, leading to the modified
version of H0 . Thus it is still open whether a similar result holds true for the resonances
of H0 . Anyhow, the imaginary part of these resonances is expected to be of order ec/ in
with c > 0 (as shown in [Ma2, Ma3] in a simpler framework). Since the inverse of the
imaginary part of a resonance is interpreted as the lifetime of the corresponding resonant
state, the resonant states in the present situation should live on a time scale of order ec/ .
Thus such a state could be mistaken for a true bound state and should perhaps be taken
into account in Chemistry, when stable molecular structures are sought.
Next we are interested in the non-resonant scattering (or collision) theory for diatomic
molecules. We use the framework presented in Section 4 for the operator H (cf. (3.13))
but we impose a stronger decay on the pair interactions. We replace at spatial infinity the
Coulomb interaction | |1 , which has long range, by a short range potential (essentially
of the type | |1 , for some > 0). To present the result in [KMW2], we need a short
review of the short range scattering theory (see details in Section XI.5 p. 75 in [RS3]).
Recall that the free dynamics for large negative time is generated by Hc in (3.19), which
can be written as
1 1
1 

Hc =
+ x + THE
+ Qc .

2 m1 m2

Here we denote by THE

+ Qc the operator on the y variables and also the same operator
acting in the x and y variables (the latter being in fact the direct integral over the x

Born-Oppenheimer approximation, 04-07-2013

20

variables of the former). We choose an eigenvalue E c and a corresponding eigenvector c

of THE
+ Qc and consider the scattering process that starts for t by the free motion
given by
1 1
1 

+
x
2 m1 m2

of two clusters that are in the state c (y). The initial state is described by a wavefunction
(x)c (y), where is a nuclear wavefunction and c (y) is actually the (tensor) product
of an electronic wavefunction of the cluster c1 by one of the cluster c2 . If we forget about

the Hughes-Eckart contribution THE

, these latter wavefunctions represent respectively an
electronic bound state in the atom/ion c1 and another in the atom/ion c2 . It turns out
that we can find a wavefunction + (x, y) such that, for t , the real evolution of
+ (x, y) is close to the free evolution of (x)c (y), that is

itH

e
+ eitHc (x)c (y) 0 .
We denote by + the operator (x)c (y) 7 + . Similarly, the final state is described
by (x )d (y ) corresponding to some cluster decomposition d (with a priori different
coordinates) and one can find a wavefunction (x , y ) such that, for t +,

itH

e
eitHd (x )d (y ) 0 .

We define ( (x )d (y )) = and we can check that (x )d (y ) can be recovered from

(x , y ) by application of the adjoint of . Thus the operator S := + sends
the initial state to the final state of the scattering process. It is the scattering operator
for this process while + and are the wave operators. The strange sign convention for

the wave operators can be interpreted in the following way: eitH + (x)c (y) represents
the future (+) evolution for the interactive dynamics (defined by H ) of the free state
(x)c (y). When c = d, = , and c = d , we have an elastic scattering process.
When c = d but c and d are orthogonal, the inelastic process corresponds to a change
of electronic level in some cluster (an excitation of an electron in c1 for instance). When
c 6= d but c1 and d1 contain the same nucleus and so do c2 and d2 , an electron at least
has moved from one nucleus to the other. We can also consider the case where c is
as above while d1 contains the two nuclei and d2 only electrons (for instance, two ions
forming a molecule and loosing some electrons). Among the inelastic processes we just
described, the two last ones might be interesting for Chemistry. The above construction
of a scattering operator can be done for all possible cluster decompositions c and d and
the collection of the S operators completely describes the possible scattering processes.
The same construction can also be performed for molecules with more than 2 nuclei with
a richer family of processes of chemical interest.
We point out that a scattering theory exists for long range interaction (like the Coulomb
one). Essentially, one has to modify the construction of the wave operators, which become
technically more involved.
We come back to the situation studied in [KMW2], that is for a diatomic molecule with
d = c given as in Section 3 but with short range interactions. We choose an energy range
(Ec , E+c ) as in fig. 2. In particular, it is above the infimum of the spectrum of Hc in
(3.19) thus, by the HVZ Theorem (see Theorem XIII.17 p. 121 in [RS4]), this energy

Born-Oppenheimer approximation, 04-07-2013

21

range is included in the continuous part of H but might contain eigenvalues. We focus
on scattering processes with total energy E (Ec , E+c ). In view of Section 4, we replace

E+ by E+c and consider eigenvalues j (x, ) of Qc (x) + THE

that are somewhere less or
c
equal to E+ . One then constructs the associated projections (x, ) and consider the
adiabatic operator H . Let c be wave operators associated to the decomposition c
as above such that the electronic energy E c of the initial state satisfies E c < Ec (E c is
close to some eigenvalue E c (0) of Qc , see fig. 2). Now one can also compare the dynamics
of H and the free dynamics generated by Hc and construct so-called adiabatic wave
operators AD
. The goal is to show that
kc AD
k = O() ,

(6.1)

when the wave operators act on wavefunctions with energy in (Ec , E+c ). To this end, one
needs an important assumption, the non-trapping condition, on the classical mechanics
generated by the nuclear, classical Hamilton functions hj (q, p) = kpk2 + j (q, 0) (with
= 0 and for the above selected eigenvalues j ) at energies in (Ec , E+c ). This nontrapping condition says that all classical trajectories of energy E (Ec , E+c ) for any
Hamilton function hj go to spatial infinity in both time directions. It implies the absence
of eigenvalues in (Ec , E+c ) and prevents resonance phenomena. In fig. 2, this assumption
is satisfied.
Under the additional assumption that only the simple eigenvalue 1 is somewhere less or
equal to E+c (in particular the image of (x, ) is always of dimension 1), the approximation
(6.1) is proved in [KMW2]. An important step in the proof is to etablish an appropriate
estimate on the resolvents C \ R z 7 (z H )1 and C \ R z 7 (z H )1 of H
and H respectively and this is done for long range interactions (in particular for the
Coulomb one). Because of the additional assumption, only elastic scattering is covered.
In this framework, we mention the papers [J2] on the scattering operator and [JKW] on
scattering cross-sections.
If one removes the above additional assumption, one can obtain the approximation (6.1)
but under the condition that the eigenvalues j do not cross (see [J1]). In this situation,
a similar approximation holds true for scattering cross-sections and it can be shown that
the inelastic scattering is disadvantaged compared to the elastic one (see [J3]). In the
simplified framework of a Schrodinger operator with matrix potential, it is even shown in
[BM] that the inelastic scattering is exponentially small in . Therefore, to study it, we
probably have to accept eigenvalues crossings and we need to control their effect on the
scattering. As mentioned before, the projection is still smooth but the eigenvalues j
might be only continuous and the corresponding eigenvectors jk might be discontinuous
at the crossing. In this situation, we mention the work by [FR] on Schrodinger operators
with matrix potential and for a special case of crossing (crossing at just one point), where
the resolvent estimates mentioned above are derived. For some types of crossing, the j
and jk are smooth and one can prove the same result (see [J4, DFJ]). This is the case
for diatomic molecules thanks to their radial symmetry with respect to the x variable. In
the work in progress [JS], one uses this property to get the resolvent estimates and also
the approximation (6.1) for diatomic molecules.
Now we come to the last field we wanted to consider, namely the Born-Oppenheimer
approximation for the time evolution of molecules, and present results obtained in [MS2].

Born-Oppenheimer approximation, 04-07-2013

22

We consider again the operator H in (3.5) but we look for an approximation of the evolution operator eitH/ (i.e. the molecular evolution on a time scale 1/). As in Section 4,
the authors choose a certain energy range (like (E ; E+ ) in fig. 1) and construct a better
projection r , starting from the operator adapted to this energy range. The estimate
of the commutator [H, ] = O() is improved in this way in the estimate [H, r ] = O(p ),
for all integer p. With the help of r , the authors introduce a map W that transforms
wavefunctions (x, y) for the full molecule into wavefunctions in x only but with values
in the L-dimensional vectors (L being the constant dimension of the image of the (x)).
This map replaces the electronic wavefunctions, that live in an infinite dimensional space,
by a finite number of degrees of freedom, namely the coordinates of the L-dimensional
vectors. We point out here that no restriction on the number of nuclei is required and
that eigenvalue crossings are allowed. There exists a L L-matrix operator A acting on
the range of W such that, for a large class of initial states with energy in the chosen
energy range, for all integer p, the time evolution of is given by
eitH/h = W eitA/h W + O((1 + |t|)p ) ,

(6.2)

where t ranges in some bounded, p- and -independent interval. So, to compute a good
approximation of the time evolution of , one first lets W act, then follow the evolution
of W generated by A (a simpler evolution) and then lets the adjoint of W act. The
operator A is obtained by an infinite but explicit construction. If one accepts an error of
size O((1 + |t|)p ), for a fixed p, one can replace A by an operator Ap which is obtained
by a finite procedure.
As a consequence of the previous approximation, the authors derive for L = 1 a rather precise description of the time evolution of coherent states (which are probably the simplest
states), completing in this way previous results of this type (for instance in [Ha1, Ha5]).
The assumption L = 1 prevents eigenvalue crossings. For the time evolution of coherent states, the effect of eigenvalue crossings was studied in [Ha7]. Even for these states,
this effect is complicated in general and another approach was followed by considering
so-called avoided crossings (see [HJ1, HJ2]). Instead of having a crossing of the electronic
eigenvalues 1 and 2 , one assumes that, for some particular nuclear position, the nonzero
difference 1 2 is small (with an appropriate size compared to ). This approach avoids
the technical difficulties carried by true crossings but allows inelastic phenomena (like the
transfer of a wave packet from the electronic level 1 to 2 ). In a simplified framework
(compared to the molecular setting) but for the time evolution through true eigenvalue
crossings, we mention [FG] in a special case where the j and the eigenvectors jk are not
smooth and [DFJ] where the latter are smooth. In [FG] a Landau-Zener formula plays
an important role. In [DFJ], although the coupling of the smooth crossing eigenvalues
vanishes formally at = 0, a coupling effect between them is proved in a very special
situation, that should be unphysical. Finally we quote the paper [TW] where the BornOppenheimer approximation for the time evolution of molecules coupled to a quantized
radiation field is analysed.
We end this section with some comment on the symmetries of particles. First one should
take into account that the electrons are fermions and consider only antisymmetric electronic wavefunctions. Second, if the molecule contains two identical nuclei for instance,

Born-Oppenheimer approximation, 04-07-2013

23

one should restrict the nuclear wavefunctions to the ones that are symmetric with respect
to the exchange of these two nuclei. In principle, such constraints can be included in
a mathematical framework but, in practice, this has not been done. Let us give some
explanation for this. Including these symmetries amounts to letting act the operators on
smaller Hilbert spaces. So if one can perform the approximation in the full Hilbert space,
it is also valid on a smaller one. However, the electronic symmetry could change the spectrum of the electronic Hamiltonian (the eigenvalue 3 could be absent or its multiplicity
could be lowered) but this would change essentially the imput of the above mathematical
treatment and not the core of the approximation. Taking into account the nuclear symmetry could give finer results but this would be hidden in the properties of the adiabatic
operator derived by the mathematical Born-Oppenheimer approximation. Up to now, it
seems that there was no clear motivation from the mathematical point of view to include
symmetries; thus it was natural to avoid them and the technical complications they carry.

Conclusion.

During the review of the traditional Born-Oppenheimer approximation (in Section 2), we
have pointed out some imprecisions that can be partially cured with the help of a sophisticated mathematical tool (see Section 5). In particular we have explained the mathematical
difficulties to derive (a precise version of) important coupled differential equations, that
are claimed to be exact in the usual, physical presentation of the theory. We have presented the way mathematicians consider the approximation and emphasise the following
main difference. In the mathematical point of view, one fixes first the total energy of the
molecular system and then constructs an effective Hamiltonian depending on this energy.
The accuracy of the Born-Oppenheimer approximation is then defined by the quality of
the replacement of the true Hamiltonian by the effective one and is measured in terms
of a small parameter related to the electron/nucluon mass ratio, enlarging in this way
the traditional definition of the sentence making the Born-Oppenheimer approximation.
This procedure turns out to be universal in the sense that it applies to different physical
situations (bound states, time-evolution, resonances, non-resonant scattering). We also
have seen that the mathematical approach actually provides a presentation of the approximation that avoids the mentioned, complicated tool and relies on more elementary
arguments and well-defined objects.
We have presented the essential structure of the mathematical justification of the BornOppenheimer approximation and tried to illustrate it on concrete results on bound states,
on the time evolution, and in scattering theory. In particular, we have seen that the basic
idea, namely the use of electronic levels and eigenfunctions, is actually the same as in
[BH, Me, SW, We] but with an important additional feature: the neglect of electronic
high energy contributions thanks to an energy argument. One writes the full Hamiltonian
as the sum of the nuclear kinetic energy, of an electronic Hamiltonian, and of comparatively smaller terms, mimicking in this way the usual framework for the well-developed
semiclassical analysis. Indeed, taking the favorite example of this analysis, namely the
semiclassical Schrodinger operator 2 x + V (x), the nuclear kinetic energy stands for
the semiclassical Laplace operator 2 x while the electronic Hamiltonian plays the role

Born-Oppenheimer approximation, 04-07-2013

24

of the potential V . Note that a semiclassical touch is already present in the original
work [BO] when expansions around a nuclear equilibrium positions are performed. We
have explained how the full Hamiltonian can be approximated by a so-called adiabatic
operator, the construction of which essentially rests upon the electronic Hamiltonian (or
clamped-nuclei Hamiltonian). Even the construction of the refined projection r , which
leads to a very accurate approximation, completely depends on this Hamiltonian. We
point out that our intuitive argument to compute an eigenvalue and an eigenvector of the
full Hamiltonian (the operator H), up to an error O(), actually leads to a modification
of Born-Huangs proposition of approximated eigenvalue and eigenvector (see (4.10)), the
change consisting in a limitation of the number of electronic levels taken into account.
This modified Born-Huangs approach is legitimate but not very accurate. To go beyond,
as we mentioned, one needs to take into account the variation of the electronic Hamiltonian with respect to the nuclear variables. When we look for an eigenvalue close to the
groundstate energy (which is close to the infimum of the lowest electronic eigenvalue 1 ),
we have seen that the nuclear kinetic energy is small, as a consequence of this closeness
and not of the large size of the nuclear masses. In particular, the original computation in
[BO] is legitimate. The situation is different for higher energy but it can be handled with
the help of semiclassical analysis (see [KMSW]), as explained in Section 6. Concerning
the scattering (or collision) theory and the time evolution of molecules, we reviewed some
results and pointed out the main difficulty, namely the control of eigenvalue crossings. In
particular, this difficulty hinders the treatment of chemically relevant situations but we
stressed that some progress was made. Letting tend to 0 instead of keeping its physical
value is essential in all the above mathematical works but, as we noticed, it might be
inappropriate in some physical or chemical situations.
In [SW], the authors subscribed to Lowdins impression (expressed in [L]), that it might
be difficult to extract from the molecular Hamiltonian the concrete realization of chemical
concepts like isomerism, conformation, chirality. Probably, they are right but the situation is perhaps not hopeless. We pointed out the papers [HJ6, HJ7, HJ8] that try to
describe hydrogen bonds. In the paper [JKW], it was proved that some symmetries in
the ion-atom scattering influence the leading term of scattering cross-sections in the large
nuclear masses limit. The techniques used in [KMSW] tells us that, near the minimum of
a nondegenerate electronic eigenvalue j , one can find a bound state of the molecule with
low nuclear kinetic energy. In this state, the nuclei vibrate near an equilibrium position,
located where the minimum is attained, in the sense that the wave function is concentrated
in the nuclear variables near this position. If one can compute (numerically) this position,
one gets the nuclear structure of this bound state (internuclear distances, symmetries).
Because of computational error, it might be difficult to check if the molecule is planar or
not. By light excitation, one can measure the difference between the molecular energies,
that are the two closest levels to the minimum of j . If the difference is very small,
then the molecule in this state is not planar and if the difference is big enough, then it
is planar, thanks to [KMSW]. Of course, these examples are limited from the chemical
point of view but show that simple properties of the molecular structure can be extracted
from the molecular Hamiltonian. We also stress that there exist tools, like the theory of
(co-)representations, to take into account symmetries of molecules. An example of such
use in the molecular context is provided in [Ha7].

Born-Oppenheimer approximation, 04-07-2013

25

We emphasise that, in the mathematical treatment of the Born-Oppenheimer approximation, the nuclei are always considered as quantum particles. The use of clamped nuclei
is just a tool to construct an appropriate effective Hamiltonian but the latter is a quantum, nuclear Hamiltonian with restricted electronic degrees of freedom. It seems that the
Born-Oppenheimer approximation in Chemistry often reduces to the computation of the
molecular potential energy surfaces and to classical motion of the nuclei on these surfaces.
Except for the special situation studied in [BO] (concerning the ground state), a classical
treatment of the nuclei is not justified, as already pointed out in [SW]. This does not mean
that classical behaviours of the nuclei do not emerge. On the contrary, it is a general fact
that semiclassical situations (like the molecular one with small ) are strongly connected
to classical features and often produce effects that resemble classical ones at the macroscopic level. For instance, we have seen that the nuclear, classical Hamiltonians hj play
a role in the scattering situation and explained above that, under some circumstances,
the nuclei can be viewed as classical particles vibrating near an equilibrium position. A
classical use of potential energy surfaces can be applied to find some molecular excited
energy near the bottom of an electronic potential well but may not capture all of them.
So we subscribe to the warning addressed to Chemists in [SW].
To the presentation of the Born-Oppenheimer approximation in [BO, BH, Me, SW, We]
we essentially added the notion of adiabatic Hamiltonian associated to a chosen total energy as a central tool (and removed some undefined objects and equations). It is present
in all the mathematical results discussed above. For this reason, the mathematicians
consider the Born-Oppenheimer approximation as valid if the true Hamiltonian can be
replaced by the adiabatic one up to some controlled error depending on the semiclassical
parameter. This property could be called an adiabatic reduction or approximation in order to avoid confusion with the traditional Born-Oppenheimer approximation in Physics.
We also insisted on the semiclassical structure of the mathematical approximation that
is close to the typical model 2 x + V (x). Our optimistic view of future mathematical
developments actually relies on the power and the diversity of techniques provided by the
semiclassical analysis.
The actual mathematical treatment of the Born-Oppenheimer approximation for molecular systems is expressed in a rather involved language and provides a theoretical information on such systems, that might be considered as unsatisfactory from the physical or
chemical point of view. We tried to make it accessible to a large readership and to show
that, despite the real difficulties it has to face, it could be improved, taking more and
more into account physical and chemical preoccupations.

Appendix

In this appendix we recall, in a intuitive but not very precise way, some mathematical
notions used in the text.
Infimum and exponential smallness. The infimum of a subset of R (the set of real numbers)
that is bounded below is the largest lower bound of this set. If this set is the image of a real
valued function, its infimum defines the infimum of the function. In the cases considered

Born-Oppenheimer approximation, 04-07-2013

26

in the text, the infimum of a function is actually its minimal value. A real or complex
valued function f of a positive is exponentially small in if there exists c, C > 0 such
that, for 0 < 1, |f ()| Cec/ . It is indeed small near 0 since Cec/ tends to 0 as
0.
L2 -spaces. The L2 -norm kf k of a complex valued function f is the integral of the modulus
squared of the function. If it is finite, we say that the function belong to the space L2
which is a normed vector space. Such a function is normalized if its L2 -norm equals one.
The L2 -norm is associated to the scalar product
Z
hf, gi = f (t)g(t) dt .
In the space L2 there are functions that are not differentiable in the usual sense. It turns
out that one can define for them generalized derivatives (in the sense of distribution). In
particular, the wave function is not everywhere regular in the usual sense but x has
a meaning as L2 function. We say that a function 0 < 7 f () L2 is O() in L2 -norm
near 0 if there exists C > 0 such that, for all 0 < 1, kf ()k C.
Operators in L2 -spaces, self-adjointness, spectrum. In the text we consider operators
acting in L2 . Such an operator A is defined on a large enough subset D(A) of the
space L2 and, for f D(A), Af is again in L2 . The range of A is the set of all Af for
f D(A). It is a vector subspace of L2 . The rank of A is the dimension of the range of
A. An operator A is bounded if there exists C > 0 such for all f D(A), kAf k Ckf k.
Such an operator can be extended to the whole L2 with the previous property preserved.
A projection P of L2 is a bounded operator defined on L2 satisfying P 2 = P (here P 2
denotes the composition of P by P ).
For appropriate operators A in L2 , one can define an adjoint denoted by A . It is defined
on an appropriate domain D(A ) and satisfies, for f D(A ) and g D(A), hA f, gi =
hf, Agi. Sometimes A is an extension of A, that is D(A) is contained in D(A ) and
Af = A f for f D(A). Such a A is symmetric. Sometimes there exist an extension B
of a symmetric A such that B = B in the sense that D(B ) = D(B) and Bf = B f for
f D(B). B is called self-adjoint. If A admits such a self-adjoint extension, one says
that A has a self-adjoint realization. One can show that the Laplace operator defined
on smooth functions with compact support is symmetric and has a (unique) self-adjoint
realization in the space of L2 functions f with (distributional) f in L2 . A self-adjoint
projection on L2 is called an orthogonal projection. It turns out that, for a self-adjoint
operator A, the operator z A for nonreal z is invertible and the inverse (z A)1 is
bounded. The function z 7 (z A)1 is called the resolvent of A. Sometimes it can be
extended to some strict subset of R. The complement of this subset is called the spectrum
(A) of A. Further informations can be found in [RS1, RS2].
Spectral resolution. A complex measure on R is a function that associated to appropriate
subsets of R a complex value. The Lebesgue measure is the oneRthat associates to these
subsets their length. One can define an integral w.r.t. : r 7 r()d(), where r()
is a complex-valued function on R. For the Lebesgue measure, this integral is the usual
integral on R. Now we come to the delicate notion of spectral resolution (see [RS1] p.
234). A spectral resolution on L2 is a projection-valued measure S on R satisfying some

27

Born-Oppenheimer approximation, 04-07-2013

properties. For appropriate subsets of R, S() is an orthogonal (self-adjoint) projection

on L2 and, if is disjoint from , then S()S( ) = S( )S() = 0. The latter property
is actually an orthogonality property. For f, g L2 , sf,g : 7 hf, S()gi is a complex
measure. The integral over R of this measure gives hf, gi. More generaly, if r is a complex
valued function of a real variable, then, for appropriate f, g L2 , one can define an
operator B satisfying
Z
Z
hf , Bgi =
r() dsf,g () =:
r() dhf , Sgi() .
R

R
In this way, one can define
R an operator-valued integral associated to S: r()dS() = B.
In particular, S() = r()dS(), where r is the characteristic
R function of . It turns
out that a self-adjoint operator A can be written as A = dS() in an essentially
unique way. The corresponding S is called the spectral resolution of A. It lives on the
spectrum (A) of A in the sense that, if does not intersect (A), then S() = 0. If E
is an eigenvalue of A (this means that there is a nonzero function in D(A) such that
A = E) then S(E) is the projection onto the space { D(A); A = E}. If E is
not an eigenvalue, then S(E) = 0. The orthogonality property mentioned above implies
in particular that, if A = E and E 6 , then, for any function f ,
h , S()f i = hS(E) , S()f i = h , S(E)S()f i = 0 .
The spectral resolution S realizes a diagonalization of A. If A has only the eigenvalues
E1 , , En in its spectrum, then S is the sum over j of the Dirac delta distribution
at EjR times the projection S(E
P j ) onto the associated spectral subspace. Furthermore
A = dS() becomes A = Ej S(Ej ) in this case, that is a diagonalization of A.
Direct integrals. Let us explain notions of direct integral associated to the Hilbert space
L2 (Rnx Rdy ) (details in a more general setting can be found in [RS4] p. 280-287). This
space may be viewed as the set L2 (Rnx ; L2 (Rdy )) of square integrable functions of x Rn
with valued in the Hilbert space L2 (Rdy ). We express this by the following direct integral
of Hilbert spaces
Z
2
n
d
2
n
2
d
L (Rx Ry ) = L (Rx ; L (Ry )) =
L2 (Rdy ) dx .
A vector L2 (Rnx Rdy ) can be written as the vector direct integral
Z
=
(x; ) dx ,

(8.1)

where (x; ) denotes the L2 (Rdy )-function y 7 (x; y). Furthermore, we recover the
L2 (Rnx Rdy )-norm kk of by the usual integral
Z
2
kk =
k(x; )k2y dx ,
where k ky denotes the L2 (Rdy )-norm. More generally, the L2 (Rnx Rdy ) scalar product
h, i is given by
Z
h, i =

h(x; ), (x; )iy dx ,

Born-Oppenheimer approximation, 04-07-2013

28

where h, iy is the L2 (Rdy ) scalar product. This construction seems to be quite artificial and
useless. (8.1) only expresses the fact that one knows completely as soon as one knows
the functions y 7 (x, y) for all x. However this contruction is used to consider parameter
dependent operators. A function x 7 A(x) with values in the self-adjoint operators in
L2y := L2 (Rdy ) and with appropriate regularity defines an operator A in L2 := L2 (Rnx Rdy )
by the requirement that, for all x and appropriate L2 , (A)(x; ) = A(x)(x; ). This
means that, for all x, the L2 function A at x is an L2y function defined by the action of
A(x) on the L2y function y 7 (x, y). One sets
Z
A =
A(x) dx .
For example, the identity operator on L2 (Rnx Rdy ) is the direct integral of the constant
function x 7 I, where I is the identity operator on L2 (Rdy ). Using this applied to a
function , we recover (8.1). In the main text, the operator acting on the full L2 is the
direct integral of the operators (x) that act on the electronic L2 space, namely L2y .
The strange integral notation may be understood in the following situation. Replace the
above Hilbert space L2 (Rnx ) by some finite dimensional space Cp . Then
Z
2
p
2
d
L (Cx ; L (Ry )) =
L2 (Rdy ) dx
is actually a finite direct sum of copies of the space L2 (Rdy ), the elements of which are
just p-uple (1 , , p ) of L2 (Rny )-functions. By passing from this finite dimensional case
to the infinite dimensional one, it is natural to replace the finite direct sum by a direct
integral.
True and generalized eigenfunctions. The operator H in the text is considered as a selfadjoint operator in L2 (in all variables). A true eigenfunction is a L2 , nonzero function
in D(H) such that, for some , H = . Since H is also a differential operator, one may
have (many) solutions of the previous equation that do not belong to L2 . Such a solution
is called a generalized eigenfunction of H. It may have a quite bad regularity.
Distributions. A distribution T on Rd is a continuous linear form on the smooth, complex
valued functions with compact support. This means that, for such functions f and g, for
complex numbers a and b, the function af + bg is still smooth and has compact support
and the complex number T (af + bg) is equal to aT (f ) + bT (g). The continuity property is
a bit involved. To a (locally) integrable function on Rd , one can associate a distribution
T defined by:
Z
f 7 T (f ) =
(x)f (x) dx .
Rd

However there are distributions that are not associated to any function. This is the case
of the Dirac distribution at x0 Rd , which is defined by f 7 f (x0 ). One can define
many operations on distributions like the sum, the multiplication by a smooth function,
the differentiation, the Fourier transform (on so-called tempered distributions) but the
convolution and especially the product are only partially defined. Here we mean that
one has to choose appropriately the two distributions entering in the product (or the
convolution).

Born-Oppenheimer approximation, 04-07-2013

29

Pseudodifferential operators. A pseudodifferential operator on L2 (Rd ) is an operator formally given by

Z
d
ei(xy) a(x, )f (y) dyd ,
f 7 Af (x) = (2)

for some appropriate function a(x, ). If a identically equals one, one see, using the Fourier
and the inverse Fourier transforms, that A is the identity operator f 7 f . If a(x, ) equals
the norm of squared then one can check in the same way that A is the Laplace operator
x . Pseudodifferential operators are a generalization of differential operators. If the
function a is matrix-valued (and f vector-valued) then A is a pseudodifferential matrix
operator.

30

Born-Oppenheimer approximation, 04-07-2013

c (Q(x ))

c (Q(x ))

3 (x )

3 (x )

eigenvalue crossing

2 (x )

2 (x )

1 (x )

1 (x )

E+
E

|x0 |
|x |
E+0
E0

inf 1

Figure 1

|x |

|x|

31

Born-Oppenheimer approximation, 04-07-2013

E+c
Ec
E c (0)
Ec

|x|
infinity

eigenvalue crossings

Figure 2

Born-Oppenheimer approximation, 04-07-2013

32

References
[ABG] W.O. Amrein, A. Boutet de Monvel, V. Georgescu: C0 -groups, commutator methods and spectral theory of N-body hamiltonians., Birkhauser 1996.
[BM] M. Benchaou, A. Martinez: Estimations exponentielles en theorie de la diffusion
pour des operateurs de Schrodinger matriciels. Ann. Inst. H. Poincare, section A, tome
71, no 6 (1999), p. 561-594. See also Numdam:
http://www.numdam.org/ .
[BH] M. Born, K. Huang: Dynamical theory of crystal lattices. Clarendon, Oxford, 1954.
[BO] M. Born, R. Oppenheimer: Zur Quantentheorie der Molekeln, Ann. Phys. 84,
(1927), 457 . A English version is available on B.T. Sutcliffes homepage at:
http://www.ulb.ac.be/cpm/people/scientists/bsutclif/main.html .
[C] J.-M. Combes: On the Born-Oppenheimer approximation. International Symposium
on Mathematical Problems in Theoretical Physics (Kyoto Univ., Kyoto, 1975), pp.
467471. Lecture Notes in Phys., 39. Springer, Berlin, 1975.
[CDS] J.-M. Combes, P. Duclos, R. Seiler: The Born-Oppenheimer approximation, in:
Rigorous atomic and molecular physics, G. Velo and A. Wightman (Eds.), Plenum
Press New-York (1981), 185212.
[CFKS] H.L. Cycon, R.G. Froese, W. Kirsch, B. Simon: Schrodinger operators with applications to quantum mechanics and global geometry. Springer, 1987.
[CS] J.-M. Combes, R. Seiler: Spectral properties of atomic and molecular systems. in
Quantum dynamics of molecules (Proc. NATO Adv. Study Inst., Univ. Cambridge,
Cambridge, 1979), pp. 435482, NATO Adv. Study Inst. Ser., Ser. B: Physics, 57,
Plenum, New York-London, 1980.
[DFJ] Th. Duyckaerts, C. Fermanian Kammerer, Th. Jecko: Degenerated codimension 1
crossing and resolvent estimates. Asymptotic Analysis, Vol. 65, N. 3-4, pp. 147-174,
2009. ArXiv: 0811.2103 , HAL: hal-00338331 .
[FG] C. Fermanian Kammerer, P. Gerard: Mesures semi-classiques et croisements de
modes. Bull. Soc. math. France 130 (2002), no 1, p. 123168.
[FR] C. Fermanian Kammerer, V. Rousse: Resolvent estimates for a Schrodinger operator
with matrix-valued potential presenting eigenvalue crossings. Application to Strichartz
estimates. Comm. in Part. Diff. Eq. 33 (2008), no. 1, p. 1944.
[FLN] S. Fujiie, C. Lasser, L. Nedelec : Semiclassical resonances for a two-level
Schrodinger operator with a conical intersection. ArXiv: 0511724 .
[Ha1] G. Hagedorn: A time-dependent Born-Oppenheimer approximation, Comm. Math.
Phys. 77 (1980), 119. See also project Euclid:
http://projecteuclid.org .

Born-Oppenheimer approximation, 04-07-2013

33

[Ha2] G. Hagedorn: High order corrections to the time-dependent Born-Oppenheimer approximation I: Smooth potentials, Ann. Math. 124, no. 3, 571590 (1986). Erratum.
Ann. Math, 126 (1987), 219.
[Ha3] G. Hagedorn: High order corrections to the time-independent Born-Oppenheimer
approximation I: Smooth potentials, Ann. Inst. H. Poincare 47 (1987), no. 1, 116. See
also Numdam:
http://www.numdam.org/ .
[Ha4] G. Hagedorn: High order corrections to the time-independent Born-Oppenheimer
approximation II: Diatomic Coulomb systems, Comm. Math. Phys. 116 (1988), no. 1,
2344. See also project Euclid:
http://projecteuclid.org .
[Ha5] G. Hagedorn: High order corrections to the time-dependent Born-Oppenheimer approximation II: Coulomb systems, Comm. Math. Phys. 117 (1988), no. 3, 2344. See
also project Euclid:
http://projecteuclid.org .
[Ha6] G. A. Hagedorn: Proof of the Landau-Zener formula in an adiabatic limit with
small eigenvalue gaps. Comm. Math. Phys. 136 (1991), no. 3, 433449. See also project
Euclid:
http://projecteuclid.org .
[Ha7] G. A. Hagedorn: Molecular propagation through electron energy level crossings.
Memoirs AMS 536, vol. 111, 1994.
[HH] G. A. Hagedorn, S. Hughes: Diatomic molecules with large angular momentum in
the Born-Oppenheimer approximation. J. Phys. A 42 (2009), no. 3, 035305, 20 pp.
[HJ1] G. Hagedorn, A. Joye: Landau-Zener transitions through small electronic eigenvalue
gaps in the Born-Oppenheimer approximation. Ann. Inst. H. Poincare Phys. Theor.
68 (1998), no. 1, 85134. See also Numdam:
http://www.numdam.org/ .
[HJ2] G. Hagedorn, A. Joye: Molecular propagation through small avoided crossings of
electron energy levels. Rev. Math. Phys. 11 (1999), no. 1, 41101.
[HJ3] G. Hagedorn, A. Joye: A Time-Dependent BornOppenheimer Approximation with
Exponentially Small Error Estimates, Comm. Math. Phys. 223, no. 3 (2001), 583626.
See also project Euclid:
http://projecteuclid.org .
[HJ4] G. Hagedorn, A. Joye: Determination of non-adiabatic scattering wave functions
in a Born-Oppenheimer model. Ann. Henri Poincare 6 (2005), no. 5, 937990. ArXiv:
0406041 .

Born-Oppenheimer approximation, 04-07-2013

34

[HJ5] G. Hagedorn, A. Joye: Mathematical analysis of Born-Oppenheimer approximations. in Spectral Theory and Mathematical Physics. A Festschrift in Honor of Barry
Simons 60th Birthday, edited by F. Gesztesy, P. Deift, C. Calvez, P. Perry, and G.W.
Schlag (Oxford Univerity Press, London, 2007), p. 203.
[HJ6] G. Hagedorn, A. Joye: A mathematical theory for vibrational levels associated with
hydrogen bonds. I. The symmetric case. Comm. Math. Phys. 274 (2007), no. 3,691715.
See also project Euclid:
http://projecteuclid.org .
[HJ7] G. Hagedorn, A. Joye: Vibrational levels associated with hydrogen bonds and semiclassical Hamiltonian normal forms. Adventures in mathematical physics, 139151,
Contemp. Math., 447, Amer. Math. Soc., Providence, RI, 2007.
[HJ8] G. Hagedorn, A. Joye: A mathematical theory for vibrational levels associated with
hydrogen bonds. II. The non-symmetric case. Rev. Math. Phys. 21 (2009), no. 2,
279313. ArXiv: 0805.4526 .
[HJ9] G. Hagedorn, A. Joye: Non-adiabatic transitions in a simple Born-Oppenheimer
scattering system. Mathematical results in quantum physics, 208212, World Sci.
Publ., Hackensack, NJ, 2011.
[HT] G. Hagedorn, J.H. Toloza: Exponentially accurate semiclassical asymptotics of lowlying eigenvalues for 2 2 matrix Schrodinger operators. J. Math. Anal. Appl. 312
(2005), no. 1, 300329.
[HRJ] G. Hagedorn, V. Rousse, S.W.J. Jilcott: The AC Stark effect, time-dependent
Born-Oppenheimer approximation, and Franck-Condon factors. Ann. Henri Poincare
7 (2006), no. 6, 10651083.
[Hu] W. Hunziker: Distortion analyticity and molecular resonance curves. Ann. Inst. H.
Poincare, section A, tome 45, no 4, p. 339-358 (1986). See also Numdam:
http://www.numdam.org/ .
[J1] Th. Jecko: Estimations de la resolvante pour une molecule diatomique dans lapproximation de Born-Oppenheimer. Comm. Math. Phys. 195, 3, 585-612, 1998. Preprint:
http://jecko.u-cergy.fr/prepublications.html .
[J2] Th. Jecko: Classical limit of elastic scattering operator of a diatomic molecule in the
Born-Oppenheimer approximation. Ann. Inst. Henri Poincare, Physique theorique, 69,
1, 1998, p. 83-131. See also Numdam:
http://www.numdam.org/ .
[J3] Th. Jecko: Approximation de Born-Oppenheimer de sections efficaces totales diatomiques. Asympt. Anal. 24 (2000), p. 1-35. Preprint:
http://jecko.u-cergy.fr/prepublications.html .

Born-Oppenheimer approximation, 04-07-2013

35

[J4] Th. Jecko: Non-trapping condition for semiclassical Schrodinger operators with
matrix-valued potentials. Math. Phys. Electronic Journal, 11 (2005), no. 2.
Erratum: Math. Phys. Electronic Journal, No. 3, vol. 13, 2007.
[J5] Th. Jecko: A new proof of the analyticity of the electronic density of molecules. .
Letters in Mathematical Physics 93, nb. 1, pp. 73-83 (2010). Arxiv: 0904.0221 ,
HAL: hal-00372603 .
[JKW] Th. Jecko, M. Klein, X.P. Wang: Existence and Born-Oppenheimer asymptotics
of the total scattering cross-section in ion-atom collisions. in Long time behaviour
of classical and quantum systems, proceedings of the Bologna AP-TEX international
conference, sept. 1999, edited by A. Martinez and S. Graffi. Preprint:
http://jecko.u-cergy.fr/prepublications.html .
[JS] Th. Jecko, V. Sordoni: Scattering of diatomic molecules and Born-Oppenheimer approximation. Work in progress.
[JK] A. Jensen, T. Kato: Spectral properties of Schrodinger operators and time decay of
wave functions. Duke Math. J. 46, 583-611, (1979).
[K] T. Kato: Pertubation theory for linear operators. Springer-Verlag 1995.
[KMSW] M. Klein, A. Martinez, R. Seiler, X.P. Wang: On the Born-Oppenheimer expansion for polyatomic molecules. Comm. Math. Phys. 143, no. 3, 607-639 (1992). See
also project Euclid:
http://projecteuclid.org .
[KMW1] M. Klein, A. Martinez, X.P. Wang: On the Born-Oppenheimer Approximation of
Wave Operators in Molecular Scattering Theory , Commun. Math. Phys. 152, (1993),
7395. See also project Euclid:
http://projecteuclid.org .
[KMW2] M. Klein, A. Martinez, X.P. Wang: On the Born-Oppenheimer Approximation of
Diatomic wave operators II. Singular potentials, Journal Math. Phys. 38 no.3, (1997),
13731396.
[L] P.-O. Lowdin: On the long way from the Coulombic Hamiltonian to the electronic
structure of molecules. Pure Appl. Chem. 61, 2065, (1989).
[Ma1] A. Martinez: Developpement asymptotiques et efffet tunnel dans lapproximation
de Born-Oppenheimer, Ann. Inst. H. Poincare 49 (1989), 239257. See also Numdam:
http://www.numdam.org/ .
[Ma2] A. Martinez: Resonances dans lapproximation de Born-Oppenheimer. I. (French)
[Resonances in the Born-Oppenheimer approximation. I] J. Differential Equations 91
(1991), no. 2, 204234.

Born-Oppenheimer approximation, 04-07-2013

36

[Ma3] A. Martinez: Resonances dans lapproximation de Born-Oppenheimer. II. Largeur

des resonances. (French) [Resonances in the Born-Oppenheimer approximation. II.
Resonance width] Comm. Math. Phys. 135 (1991), no. 3, 517530. See also project
Euclid:
http://projecteuclid.org .
[Ma4] A. Martinez: Eigenvalues and resonances of polyatomic molecules in the BornOppenheimer approximation. Schrodinger operators (Aarhus, 1991), 145152, Lecture
Notes in Phys., 403, Springer, Berlin, 1992.
[MM] A. Martinez, B. Messerdi: Resonances of diatomic molecules in the BornOppenheimer approximation, Comm. Part. Diff. Eq. 19 (1994), 11391162.
[MS1] A. Martinez, V. Sordoni: A general reduction scheme for the time-dependent BornOppenheimer approximation. C. R. Math. Acad. Sci. Paris 334 (2002), no. 3, 185188.
[MS2] A. Martinez, V. Sordoni: Twisted pseudodifferential calculus and application to the
quantum evolution of molecules. Memoirs Am. Math. Soc., Vol. 200, n. 936 (2009).
ArXiv: 0809.3663 .
[Me] A. Messiah: Quantum Mechanics, (1960), Vol. II. North-Holland, Amsterdam.
[N] L. Nedelec: Resonances for matrix Schrodinger operators. Duke Math. J. 106 (2001),
no. 2, p. 209-236.
[PST] G. Panati, H. Spohn, S. Teufel: The time-dependent Born-Oppenheimer approximation, ESIAM: Math. Model. and Num. Anal. 41 (2007), 297314. ArXiv:
0712.4369 .
[Ra] A. Raphaelian: Ion-atom scattering within a Born-Oppenheimer framework, Dissertation Technische Universitat Berlin (1986).
[RS1] M. Reed, B. Simon: Methods of Modern Mathematical Physics, Vol. I: Functional
Analysis. Academic Press, 1975.
[RS2] M. Reed, B. Simon: Methods of Modern Mathematical Physics, Vol. II: Fourier
Analysis, Self-adjointness. Academic Press, 1975.
[RS3] M. Reed, B. Simon: Methods of Modern Mathematical Physics, Vol. III: Scattering
theory. Academic Press, 1979.
[RS4] M. Reed, B. Simon: Methods of Modern Mathematical Physics, Vol. IV: Analysis
of operators. Academic Press, 1978.
[R] V. Rousse: Landau-Zener transitions for eigenvalue avoided crossings in the adiabatic
and Born-Oppenheimer approximations. Asymptot. Anal. 37 (2004), no. 3-4, 293328.
[ST] H. Spohn, S. Teufel: Adiabatic decoupling and time-dependent Born-Oppenheimer
theory, Comm. Math. Phys. 224, no. 1, (2001) 113132. ArXiv: 0104024 .

Born-Oppenheimer approximation, 04-07-2013

37

[SW] B.T. Sutcliffe, R.G. Woolley: On the quantum theory of molecules. J. Chemical
Physics 137, 22A544 (2012). ArXiv: 1206.4239 .
[TW] S. Teufel, J. Wachsmuth: Spontaneous decay of resonant energy levels for molecules
with moving nuclei., Comm. Math. Phys. 315, no. 3, (2012) 699738. ArXiv:
1109.0447 .
[Wa1] X.P. Wang: Asymptotic behaviour of resolvent of N-body Schrodinger operators
near a threshold., Ann. H. Poincare 4 (2003) 553-600. Preprint available on:
http://www.math.sciences.univ-nantes.fr/ wang/ .
[Wa2] X.P. Wang: On the existence of the N-body Efimov effect., J. Funct. Anal. 209
(2004) 137-161.
Preprint available on:
http://www.math.sciences.univ-nantes.fr/ wang/ .
[We] S. Weinberg: Lectures on Quantum Mechanics., Cambridge University Press
(November 30, 2012).
[Y] D.R. Yafaev: On the theory of the discrete spectum of the three-particle Schr
odinger
operator., Math. USSR-Sb 23, (1974), 535-559. (November 30, 2012).