Beruflich Dokumente
Kultur Dokumente
Simen Kvaal*
Centre of Mathematics for Applications, University of Oslo, N-0316 Oslo, Norway
共Received 15 August 2008; revised manuscript received 15 May 2009; published 27 July 2009兲
We give a thorough analysis of the convergence properties of the configuration-interaction method as applied
to parabolic quantum dots among other systems, including a priori error estimates. The method converges
slowly in general, and in order to overcome this, we propose to use an effective two-body interaction well
known from nuclear physics. Through numerical experiments we demonstrate a significant increase in accu-
racy of the configuration-interaction method.
II. HARMONIC OSCILLATOR AND PARABOLIC The eigenvalue of n共x兲 is n + 1 / 2 so that the eigenvalue
QUANTUM DOTS of ⌽␣共rជ兲 is
A. Harmonic Oscillator d
⑀␣ = + 兩␣兩, 共2.10兲
A spinless particle of mass m in an isotropic harmonic 2
potential has Hamiltonian
i.e., a zero-point energy d / 2 plus a non-negative integer. We
ប2 2 1 denote by 兩␣兩 the shell number of ⌽␣ and the eigenspace
HHO = − ⵜ + m2储rជ储2 , 共2.1兲
2m 2 Sr共Rd兲 corresponding to the eigenvalue d / 2 + r a shell. We
define the shell-truncated Hilbert space PR共Rd兲 傺 L2共Rd兲 as
where rជ 苸 Rd is the particle’s coordinates. By choosing
proper energy and length units, i.e., ប and 冑ប / m, respec- R
冉 冊
d
1 2 1 2 Since the Hermite functions constitute an orthonormal ba-
HHO = 兺 − + r , 共2.3兲 sis for L2共Rd兲, PR共Rd兲 → L2共Rd兲, in the sense that for every
2 r2k 2 k
k=1 苸 L2共Rd兲, limR→⬁储 − P储 = 0, where P is the orthogonal
so that a complete specification of the HO eigenfunctions is projector on PR共Rd兲. Strictly speaking, we should use a sym-
given by bol such as PR or even PR共Rd兲 for the projector. However, R
and d will always be clear from the context, so we are de-
⌽␣1,␣2,. . .,␣d共rជ兲 = ␣1共r1兲␣2共r2兲 ¯ ␣d共rd兲, 共2.4兲 liberately sloppy to obtain a concise formulation. For the
same reason, we will sometimes simply write P or PR for the
where ␣i共x兲, ␣i = 0 , 1 , . . . are one-dimensional HO eigen-
space PR共Rd兲.
functions, also called Hermite functions. These are defined An important fact is that since HHO is invariant under
by orthogonal spatial transformations 共i.e., such transformation
2
n共x兲 = 共2nn ! 1/2兲−1/2Hn共x兲e−x /2, n = 0,1, . . . , conserve energy兲 so is each individual shell space. Hence,
each shell Sr共Rd兲, and also PR共Rd兲, is independent of the
共2.5兲 spatial coordinates chosen.
where the Hermite polynomials Hn共x兲 are given by For the case d = 1 each shell r is spanned by a single
eigenfunction, namely, r共x兲. For d = 2, each shell r has de-
2 n −x2 generacy r + 1, with eigenfunctions
Hn共x兲 = 共− 1兲nex e . 共2.6兲
xn
⌽共s,r−s兲共rជ兲 = s共r1兲r−s共r2兲, 0 ⱕ s ⱕ r. 共2.12兲
The Hermite polynomials also obey the recurrence formula
The usual HO eigenfunctions used to construct many-
Hn+1共x兲 = 2xHn共x兲 − 2nHn−1共x兲, 共2.7兲 body wave functions are not the Hermite functions ⌽␣1,¯,␣d,
with H0共x兲 = 1 and H1共x兲 = 2x. The Hermite polynomial Hn共x兲 however, but rather those obtained by utilizing the spherical
has n zeroes, and the Gaussian factor in n共x兲 will eventually symmetry of the HO. This gives a many-body basis diagonal
subvert the polynomial for large 兩x兩. Thus, qualitatively, the in angular momentum. For d = 2 we obtain the so-called
Hermite functions can be described as localized oscillations Fock-Darwin orbitals given by
with n nodes and a Gaussian “tail” as x approaches ⫾⬁. One
can easily compute the quantum-mechanical variance ⌽n,m
FD
共r, 兲 = 冋 2n!
共n + 兩m兩兲!
册 1/2 im
e
冑2 Ln
兩m兩 2
共r2兲e−r /2 . 共2.13兲
共⌬x兲2 ª 冕−⬁
⬁
1
x2n共x兲2dx = n + ,
2
共2.8兲 Here, n ⱖ 0 is the nodal quantum number, counting the nodes
of the radial part, and m is the azimuthal quantum number.
showing that, loosely speaking, the width of the oscillatory The eigenvalues are
region increases as 共n + 1 / 2兲1/2.
The functions ⌽␣1,. . .,␣d defined in Eq. 共2.4兲 are called ⑀n,m = 2n + 兩m兩 + 1. 共2.14兲
d-dimensional Hermite functions. In the sequel, we will de- Thus, R = 2n + 兩m兩 is the shell number. By construction, the
fine ␣ = 共␣1 , . . . , ␣d兲 苸 Id for a tuple of non-negative integers, Fock-Darwin orbitals are eigenfunctions of the angular-
also called a multi-index; see the Appendix. Using multi- momentum operator Lz = −i / with eigenvalue m. Of
indices, we may write course, we may write ⌽n,m FD
as a linear combination of the
2 Hermite functions ⌽共s,R−s兲, where 0 ⱕ s ⱕ R = 2n + 兩m兩, and
⌽␣共rជ兲 = 共2兩␣兩␣ ! d/2兲−1/2H␣1共r1兲 ¯ H␣d共rd兲e−储rជ储 /2 .
vice versa. The actual choice of form of eigenfunctions is
共2.9兲 immaterial as long as we may identify those belonging to a
045321-2
HARMONIC OSCILLATOR EIGENFUNCTION EXPANSIONS,… PHYSICAL REVIEW B 80, 045321 共2009兲
n=0 n=0
m = −1 m=1 α = (0, 1) α = (1, 0)
n=0
m=0 α = (0, 0)
FIG. 1. Illustration of PR=4共R2兲: 共left兲 Fock-Darwin orbitals. 共right兲 Hermite functions. Basis functions with equal HO energy are shown
at same line.
given shell. The space PR=4共R2兲 is illustrated in Fig. 1 using well understood.19–21 Here, we consider d = 2 dimensions
both Hermite functions and Fock-Darwin orbitals. only, but the d = 3 case is similar. We note that for N = 2 it is
enough to study the spatial wave function since it must be
B. Parabolic quantum dots either symmetric 共for the singlet S = 0 spin state兲 or antisym-
metric 共for the triplet S = 1 spin states兲. Hamiltonian 共2.15兲
We consider N electrons confined in a harmonic oscillator becomes
in d dimensions. This is a very common model for a quan-
tum dot. We comment that modeling the quantum dot geom- 1 1
H = − 共ⵜ21 + ⵜ22兲 + 共r21 + r22兲 + , 共2.19兲
etry by a perturbed harmonic oscillator is justified by self- 2 2 r12
consistent calculations13–15 and is a widely adopted
assumption.5,8,9,16–18 where r12 = 储rជ1 − rជ2储 and r j = 储rជ j储. Introduce a set of scaled
The Hamiltonian of the quantum dot is given by center-of-mass coordinates given by R ជ = 共rជ1 + rជ2兲 / 冑2 and rជ
= 共rជ1 − rជ2兲 / 冑2. This coordinate change is orthogonal and sym-
H ª T + U, 共2.15兲 metric in R4. This leads to the separable Hamiltonian
where T is the many-body HO Hamiltonian, given by 1 1
ជ 储 2兲 +
H = − 共ⵜr2 + ⵜR2 兲 + 共储rជ储2 + 储R
N 2 2 冑2储rជ储
T = 兺 HHO共rជk兲 共2.16兲
k=1 ជ 兲 + Hrel共rជ兲.
= HHO共R
and U is the interelectron Coulomb interactions. In dimen- A complete set of eigenfunctions of H can now be written on
sionless units the interaction is given by, product form, viz.,
quantum dot then approaches the classical regime.1 where 兩m兩 ⱖ 0 is an integer and un,m is an eigenfunction of the
radial Hamiltonian given by
C. Exact solution for two electrons
1 兩m兩2 1 2
Hr = − r + 2 + r + 共2.23兲
Before we discuss the approximation properties of the 2r r r 2r 2 冑2r .
Hermite functions, it is instructive to consider the very sim-
plest example of a two-electron parabolic quantum dot and By convention, n counts the nodes away from r = 0 of un,m共r兲.
the properties of the eigenfunctions since this case admits Moreover, odd 共even兲 m gives antisymmetric 共symmetric兲
analytical solutions for special values of and is otherwise wave functions ⌿共rជ1 , rជ2兲. For any given 兩m兩, it is quite easy
045321-3
SIMEN KVAAL PHYSICAL REVIEW B 80, 045321 共2009兲
to deduce that the special value = 冑2兩m兩 + 1 yields the eigen- 兩cn兩 = o共n−共k+1+⑀兲/2兲. 共2.28兲
function
Here, k is the number of times ⌿ may be differentiated
2
u0,m = Dr兩m兩共a + r兲e−r /2 , 共2.24兲 weakly, i.e., ⌿ 苸 Hk共R2兲, and ⑀ 苸 关0 , 1兲 is a constant. For the
function ⌿0 we have k = 1. This kind of estimate directly tells
where D and a are constants. The corresponding eigenvalue us that an approximation using only a few HO eigenfunc-
of Hr is Er = 兩m兩 + 2, and E = 2n⬘ + 兩m⬘兩 + 1 + Er. Thus, the tions necessarily will give an error depending directly on the
ground state 共having m = m⬘ = 0 and n = n⬘ = 0兲 for = 1 is smoothness k.
given by We comment that for higher 兩m兩 the eigenstates will still
have cusps, albeit in the higher derivatives.22 Indeed, we
ជ ,rជ兲 = D共r + a兲e−共r2+R2兲/2 =
⌿0共R
D
冑2a兲e−共r21+r22兲/2 ,
冑2 共r12 + have weak derivatives of order 兩m兩 + 1, as can easily be de-
duced by operating on 0,m with x and y. Moreover, recall
with D being a 共new兲 normalization constant. that 兩m兩 = 1 is the S = 1 ground state, which then will have
Observe that this function has a cusp at r = 0, i.e., at the coefficients decaying faster than the S = 0 ground state. More-
origin x = y = 0 关where we have introduced Cartesian coordi- over, there will be excited states, i.e., states with 兩m兩 ⬎ 1, that
nates rជ = 共x , y兲 for the relative coordinate兴. Indeed, the partial also have more quickly decaying coefficients 兩cn兩. This will
derivatives x0,0 and y0,0 are not continuous there, and ⌿0 be demonstrated numerically in Sec. V.
has no partial derivatives 共in the distributional sense, see the In fact, Hoffmann-Ostenhof et al.22 showed that near r12
Appendix兲 of second order. The cusp stems from the famous = 0, for arbitrary any local solution ⌿ of 共H − E兲⌿ = 0 has
“cusp condition,” which in simple terms states that, for a the form
nonvanishing wave function at r12 = 0, the Coulomb diver-
gence must be compensated by a similar divergence in the
Laplacian.22,23 This is only possible if the wave function has
⌿共兲 = 储储m P 冉 冊
储储
共1 + a储储兲 + O共储储m+1兲, 共2.29兲
⌽n,0
FD
共r兲 = 冑 2
2
Ln共r2兲e−r /2 , 共2.25兲
SERIES
A. Hermite functions in one dimension
using the fact that these are independent of . Thus, In this section, we consider some formal mathematical
propositions whose proofs are given in the Appendix, and
⬁
discuss their importance for expansions in HO basis func-
⌿0共rជ兲 = ⌽0,0
FD
共R兲u0,0共r兲 = ⌽0,0
FD
共R兲 兺
n=0
cn⌽n,0
FD
共r兲, 共2.26兲 tions.
The first proposition considers the one-dimensional case,
The functions ⌽n,0 FD
共r兲 are very smooth, as is seen by noting and the second considers general multidimensional expan-
that Ln共r 兲 = Ln共x2 + y 2兲 is a polynomial in x and y, while
2 sions. The treatment for one-dimensional Hermite functions
u0,0共r兲 = u0,0共冑x2 + y 2兲, so Eq. 共2.26兲 is basically approximat- is similar, but not equivalent to, that given by Boyd25 and
ing a square root with a polynomial. Hille.26
Consider then a truncated expansion ⌿0,R 苸 PR共R2兲, such We stress that the results are valid for any given wave
as the one obtained with the CI or coupled cluster method.24 function—not only eigenfunctions of quantum dot
In general, this is different from PR⌿0, which is the best Hamiltonians—assuming only that the wave function decays
approximation of the wave function in PR共R2兲. In any case, exponentially as 兩x兩 → ⬁. In the Appendix, more general con-
this expansion, consisting of the R + 1 terms such as those of ditions are also considered.
Eq. 共2.26兲 is a very smooth function. Therefore, the cusp at The results are stated in terms of weak differentiability of
r = 0 cannot be well approximated. the wave function, which is a generalization of the classical
In Sec. III C, we will show that the smoothness properties notion of a derivative. The space Hk共R兲 傺 L2共R兲 is roughly
of the wave function ⌿ is equivalent to a certain decay rate defined as the 共square-integrable兲 functions 共x兲 having k
of the coefficients cn in Eq. 共2.26兲 as n → ⬁. In this case, we 共square-integrable兲 derivatives mx 共x兲 and 0 ⱕ m ⱕ k. Corre-
will show that spondingly, the space Hk共Rn兲 傺 L2共Rn兲 consists of the func-
tions whose partial derivatives of total order ⱕk are square
⬁
integrable. For wave functions of electronic systems, it turns
兺 nk兩cn兩2 ⬍ + ⬁
n=0
共2.27兲 out that k times continuous differentiability implies k + 1
times weak differentiability.22 The order k of differentiability
so that is not always known, but an upper or lower bound can often
045321-4
HARMONIC OSCILLATOR EIGENFUNCTION EXPANSIONS,… PHYSICAL REVIEW B 80, 045321 共2009兲
冉兺 冊
共x兲 = 兺 cnn共x兲, 共3.1兲 ⬁ 1/2
n=0
储共1 − P兲储 = p共r兲 . 共3.10兲
where n共x兲 is given by Eq. 共2.5兲. Then 苸 Hk共R兲 if and r=R+1
only if In applications, we often observe a decay of nonintegral
⬁ order; i.e., there exists an ⑀ 苸 关0 , 1兲 such that we observe
兺
n=0
nk兩cn兩2 ⬍ ⬁. 共3.2兲 p共r兲 = o共r−共k+1+⑀兲兲. 共3.11兲
This does not, of course, contradict the results. To see this,
We notice that the latter implies that
we observe that if 苸 Hk共Rn兲 but 苸 Hk+1共Rn兲, then p共r兲
兩cn兩 = o共n−共k+1兲/2兲, 共3.3兲 must decay at least as fast as o共r−共k+1兲兲 but not as fast as
o共r−k+2兲. Thus, the actual decay exponent can be anything
which shows that the more 共x兲 can be differentiated, the inside the interval 关k + 1 , k + 2兲.
faster the coefficients will fall off as n → ⬁. Moreover, let Consider also the case where 苸 Hk共Rn兲 for every k, i.e.,
R = PR = 兺n=0
R
cnn. Then we can differentiate it 共weakly兲 as many times we like. Then
冉兺 冊 ⬁ 1/2
p共r兲 decays faster than r−共k+1兲, for any k ⱖ 0, giving so-called
exponential convergence of the Hermite series. Hence, func-
储 − R储 = 兩cn兩 2
, 共3.4兲 tions that are best approximated by Hermite series are rap-
n=R+1
idly decaying and very smooth functions . This would be
which gives an estimate of how well a finite basis of Hermite the case for the quantum dot eigenfunctions if the interpar-
functions will approximate 共x兲 in the norm. We already ticle interactions were nonsingular.
notice that for low k = 2, which is typical, the coefficients fall
off as o共n−3/2兲, which is rather slowly. B. Many-body wave functions
In the general n-dimensional case, the wave function
苸 L2共Rn兲 has an expansion in the n-dimensional Hermite We now discuss N-body eigenfunctions of the HO in d
functions ⌽␣共x兲 and ␣ 苸 In given by dimensions, including spin, showing that we may identify
the expansion of such with 2N expansions in Hermite func-
共x兲 = 兺 c␣⌽␣共x兲 = 兺 c␣1¯␣n␣1共x1兲 ¯ ␣n共xn兲. tions in n = Nd dimensions, i.e., 2N expansions in HO eigen-
␣ ␣1¯␣n functions of imagined spinless particles in n = Nd dimen-
sions. Each expansion corresponds to a different spin
In order to obtain useful estimates on the error, we need to configuration.
define the shell weight p共R兲 by the overlap of 共x兲 with the Each particle k = 1 , . . . , N has both spatial degrees of free-
single shell SR, i.e., dom rជk 苸 Rd and a spin coordinate k 苸 兵⫾1其, corresponding
to the z-projection Sz = ⫾ ប2 of the electron spin. The configu-
p共R兲 = 储P共SR兲储2 = 兺 兩c␣兩2 , 共3.5兲 ration space can thus be taken as two copies X of Rd; one for
␣,兩␣兩=R
each spin value, i.e., X = Rd ⫻ 兵⫾1其 and xk = 共rជk , k兲 苸 X are
where P共SR兲 is the projection onto the shell. Thus, the coordinates of particle k.
For a single particle with spin, the Hilbert space is now
⬁ L2共X兲, with basis functions given by
储储2 = 兺 p共R兲. 共3.6兲
R=0 ⌽̂i共x兲 = ⌽␣共rជ兲共兲, 共3.12兲
For the one-dimensional case, we of course have p共R兲 where i = i共␣ , 兲 is a new generic index and where is a
= 兩cR兩2. basis function for the spinor space C2.
Proposition 2 共Approximation in n dimensions兲. Ignoring the Pauli principle for the moment, the N-body
Let 苸 L2共Rn兲 be exponentially decaying as 储x储 → ⬁ and Hilbert space is now given by
given by
H共N兲 = L2共X兲N ⬅ L2共RNd兲 丢 共C2兲N , 共3.13兲
共x兲 = 兺 c␣⌽␣共x兲. 共3.7兲
␣
i.e., each wave function 苸 H共N兲 is equivalent to 2N spin-
component functions 共兲 苸 L2共RNd兲 and = 共1 , . . . , N兲
Then 苸 Hk共Rn兲 if and only if 苸 兵⫾1其N. We have
045321-5
SIMEN KVAAL PHYSICAL REVIEW B 80, 045321 共2009兲
where = 共rជ1 , . . . , rជN兲, and where  = 共␣1 , . . . , ␣N兲 is an We return to the exact solutions of the two-electron quan-
Nd-component multi-index. Correspondingly, a basis for the tum dot considered in Sec. II C. Recall, that the wave func-
complete space H共N兲 = L2共X兲N is given by the functions tions were on the form
再
PAS,R = span ⌿i1,. . .,iN:i1 = i ⬍ ¯ ⬍ iN, 兺k 兩␣k兩 ⱕ R 冎 , Let us mention some results, mainly due to Hoffmann-
Ostenhof et al.,22,28 concerning smoothness of many-electron
共3.19兲 wave functions. Strictly speaking, their results are valid only
in d = 3 spatial dimensions since the Coulomb interaction in
which is precisely the computational basis used in many CI d = 2 dimensions fails to be a Kato potential, the definition of
calculations. 共See however also the discussion in Sec. IV.兲 which is quite subtle and out of the scope for this article.28
We stress that PAS,R is independent of the actual one-body On the other hand, it is reasonable to assume that the results
HO eigenfunctions used. The shell weight of 苸 HAS共X兲 is will still hold true since the analytical results of the N = 2
now given by case is very similar in the d = 2 and d = 3 cases: the eigen-
functions decay exponentially with the same cusp singulari-
p共r兲 = 兺
i1¯iN
具⌿i1¯iN, 典␦兩共i1¯iN兲兩,r , 共3.20兲 ties at the origin.19,20
Consider the Schrödinger equation 共H − E兲共兲 = 0, where
and = 共1 , ¯ , Nd兲 = 共rជ1 , ¯ , rជN兲 苸 RNd and where 共兲 is only
045321-6
HARMONIC OSCILLATOR EIGENFUNCTION EXPANSIONS,… PHYSICAL REVIEW B 80, 045321 共2009兲
assumed to be a solution locally. 共A proper solution is of instead of the global shell number 共or energy兲 is also
course also a local solution.兲 Recall that has 2N spin com- common.6,8 For obvious reasons, PR共N兲 is often referred to
ponents 共兲. Define a coalesce point CP as a point where at as an “energy cut space,” while MR共N兲 is referred to as a
least two particles coincide, i.e., rជ j = rជᐉ and j ⫽ ᐉ. Away from “direct product space.”
the set of such points, 共兲共兲 is real analytic since the inter- As in Sec. III A, PR is the orthogonal projector onto the
action is real analytic there. Near a CP, the wave function model space PR共N兲. We also define QR = 1 − PR as the projec-
has the form tor onto the excluded space PR共N兲⬜. The discrete eigenvalue
problem is then
共兲共 + CP兲 = rk P共/r兲共1 + ar兲 + O共rk+1兲, 共3.28兲
共PRHPR兲h,k = Eh,kh,k, k = 1, . . . ,kmax . 共4.3兲
where r = 储储, P is a hyperspherical harmonic 共on the sphere
SNd−1兲 of degree k = k共CP兲, and where a is a constant. It is The CI method becomes, in principle, exact as R → ⬁.
immediately clear that 共兲共兲 is k + 1 times weakly differen- Indeed, a widely used name for the CI method is “exact
tiable in a neighborhood of CP. However, at K-electron coa- diagonalization,” being somewhat a misnomer as only a very
lesce points, i.e., at points CP where K different electrons limited number of degrees of freedom per particle are
coincide, the integer k may differ. Using exponential decay achievable.
of a proper eigenfunction, we have 共兲 苸 Hmin共k兲+1共RNd兲. It is clear that
Hoffmann-Ostenhof et al.22,28 also showed that symmetry re-
strictions on the spin components due to the Pauli principle PR共N兲 傺 MR共N兲 傺 PNR共N兲 共4.4兲
induces an increasing degree k of the hyperspherical har- so that studying the convergence in terms of PR共N兲 is suffi-
monic P, generating even higher order of smoothness. A gen- cient. In our numerical experiments we therefore focus on
eral feature is that the smoothness increases with the number the energy cut model space. A comparison between the con-
of particles. vergence of the two spaces is, on the other hand, an interest-
However, their results in this direction are not general ing topic for future research.
enough to ascertain the minimum of the values for k for a Using the results in Refs. 30 and 33 for nondegenerate
given wave function, although we feel rather sure that such eigenvalues for simplicity, we obtain an estimate for the error
an analysis is possible. Suffice it to say that the results are in the numerical eigenvalue Eh as
clearly visible in the numerical calculations in Sec. V.
Another interesting direction of research has been under- Eh − E ⱕ 关1 + 共R兲兴共1 + K兲具,QRT典, 共4.5兲
taken by Yserentant,29 who showed that there are some very
high order mixed partial derivatives at coalesce points. It where K is a constant, and where 共R兲 → 0 as R → ⬁. Using
seems unclear, though, if this can be exploited to improve the T⌽ = 共Nd / 2 + 兩兩兲⌽ and Eq. 共3.8兲, we obtain
冉 冊
CI calculations further. ⬁
Nd
具,QRT典 = 兺
r=R+1 2
+ r p共r兲. 共4.6兲
IV. CONFIGURATION-INTERACTION METHOD
Assume now, that 共兲 苸 Hk共RNd兲 for all so that according
A. Convergence analysis using HO eigenfunction basis to proposition 2, we will have
The basic problem is to determine a few eigenvalues and ⬁
eigenfunctions of the Hamiltonian H in Eq. 共2.15兲, i.e.,
兺 rk p共r兲 ⬍ + ⬁ 共4.7兲
H k = E k k, k = 1, . . . ,kmax . 共4.1兲 r=0
The CI method consists of approximating eigenvalues of H implying that rp共r兲 = o共r−k兲. We then obtain, for k ⬎ 1,
with those obtained by projecting the problem onto a finite- 具,共1 − PR兲T典 = o共R−共k−1兲兲 + o共R−k兲. 共4.8兲
dimensional subspace Hh 傺 H共N兲. As such, it is an example
of the Ritz-Galerkin variational method.30,31 We comment For k = 1 共which is the worst case兲, we merely obtain conver-
that the convergence of the Ritz-Galerkin method is not sim- gence, 具 , 共1 − PR兲T典 → 0 as R → ⬁. We assume, that R is
ply a consequence of the completeness of the basis sufficiently large, so that the o共R−k兲 term can be neglected.
functions.30 We will analyze the CI method when the model Again, we may observe a slight deviation from the decay,
space is given by and we expect to observe eigenvalue errors on the form
再
Hh = PRHAS共N兲 = PR共N兲 = span ⌿i1,. . .,iN: 兺 兩␣k兩 ⱕ R ,
k
冎 where 0 ⱕ ⑀ ⬍ 1.
Eh − E ⬃ 共1 + K兲R−共k−1+⑀兲 , 共4.9兲
used in Refs. 10 and 32, for example, although other spaces As for the eigenvector error 储h − 储 共recall that h
also are common. 共We drop the subscript “AS” from now ⫽ PR兲, we mention that
on.兲 The space
储h − 储 ⱕ 关1 + 共R兲兴关共1 + K兲具,共1 − PR兲T典兴1/2 ,
MR共N兲 ª span兵⌿i1,. . .,iN:max兩␣k兩 ⱕ R其, 共4.2兲
k 共4.10兲
i.e., a cut in the single-particle shell numbers 共or energy兲 where 共R兲 → 0 as R → ⬁.
045321-7
SIMEN KVAAL PHYSICAL REVIEW B 80, 045321 共2009兲
B. Effective interaction scheme The effective two-body interaction Ceff共i , j兲 is now given
Effective interactions have a long tradition in nuclear by
physics, where the bare nuclear interaction is basically un- Ceff共1,2兲 ª Heff − PRTPR , 共4.15兲
known and highly singular and where it must be renormal-
ized and fitted to experimental data.3 In quantum chemistry which is defined only within PR共N = 2兲.
and atomic physics, the Coulomb interaction is of course The N-body effective Hamiltonian is defined by
well known so there is no intrinsic need to formulate an N
effective interaction. However, in lieu of the in general low Heff ª PRTPR + 兺 Ceff共i, j兲, 共4.16兲
order of convergence implied by Eq. 共4.9兲, we believe that i⬍j
HO-based calculations such as the CI method in general may
benefit from the use of effective interactions. where PR projects onto PR共N兲, and thus Heff is defined only
A complete account of the effective interaction scheme within PR共N兲. The diagonalization of Heff共N兲 is equivalent to
outlined here is out of scope for the present article, but we a perturbation technique where a certain class of diagrams is
refer to Refs. 2, 11, and 34–36 for details as well as numeri- summed to infinite order in the full problem.34 In implemen-
cal algorithms. tations, Eqs. 共4.12兲 and 共4.16兲 are treated in COM coordi-
Recall, that the interaction is given by nates, utilizing block diagonality of both H and Heff; see Ref.
36 for details.
N N
We comment that unlike the bare Coulomb interaction,
U = 兺 C共i, j兲 = 兺 , 共4.11兲 the effective two-body interaction Ceff corresponds to a non-
i⬍j i⬍j 储rជi − rជ j储
local potential due to the “straightening out” of truncated
a sum of fundamental two-body interactions. For the N = 2 eigenvectors. Rigorous mathematical treatment of the con-
problem we have in principle the exact solution since Hamil- vergence properties of the effective interaction is, to the au-
tonian 共2.19兲 can be reduced to a one-dimensional radial thor’s knowledge, not available. Effective interactions have,
equation, e.g., the eigenproblem of Hr defined in Eq. 共2.21兲. however, enjoyed great success in the nuclear physics com-
This equation may be solved to arbitrarily high precision munity, and we strongly believe that we soon will see suffi-
using various methods, for example using a basis expansion cient proof of the improved accuracy with this method. In-
in generalized half-range Hermite functions.37 In nuclear deed, in Sec. V we see clear evidence of the accuracy boost
physics, a common approach is to take the best two-body CI when using an effective interaction.
calculations available, where R = O共103兲, as “exact” for this
purpose.
We now define the effective Hamiltonian for N = 2 as a V. NUMERICAL RESULTS
Hermitian operator Heff defined only within PR共N = 2兲 that
A. Code description
gives K = dim关PR共N = 2兲兴 exact eigenvalues Ek of H, and K
approximate eigenvectors eff,k. Of course, there are infi- We now present numerical results using the full
nitely many choices for the K eigenpairs, but by treating U configuration-interaction method for N = 2 – 5 electrons in d
= / r12 as a perturbation, and “following” the unperturbed = 2 dimensions. We will use both the “bare” Hamiltonian
HO eigenpairs 共 = 0兲 through increasing values of , one H = T + U and effective Hamiltonian 共4.16兲.
makes the eigenvalues unique.2,38 The approximate eigen- Since the Hamiltonian commutes with angular momentum
vectors eff,k 苸 PR共N = 2兲 are chosen by minimizing the dis- Lz, the latter taking on eigenvalues M 苸 Z, the Hamiltonian
tance to the exact eigenvectors k 苸 H共N = 2兲 while retaining matrix is block diagonal. 共Recall that the Fock-Darwin orbit-
orthonormality.35 This uniquely defines Heff for the two-body als ⌽n,m
FD
are eigenstates of Lz with eigenvalue m, so each
system. In terms of matrices, we have Slater determinant has eigenvalue M = 兺k=1 N
mk.兲 Moreover,
the calculations are done in a basis of joint eigenfunctions
Heff = Ũdiag共E1, . . . ,EK兲Ũ† , 共4.12兲 for total electron spin S2 and its projection Sz as opposed to
the Slater determinant basis used for convergence analysis.
where X and Y are unitary matrices defined as follows. Let U
Such basis functions are simply linear combinations of Slater
be the K ⫻ K matrix whose kth column is the coefficients of
determinants within the same shell and further reduce the
PRk. Then the singular value decomposition of U can be
dimensionality of the Hamiltonian matrix.8 The eigenfunc-
written as
tions of H are thus labeled with the total spin S = 0 , 1 , . . . , N2
U = X ⌺ Y†, 共4.13兲 for even N and S = 21 , 23 , . . . , N2 for odd N, as well as the total
angular momentum M = 0 , 1 , . . .. 共−M produce the same ei-
where 兺 is diagonal. Then,
genvalues as M, by symmetry.兲 We thus split PR共N兲 共or
Ũ ª XY † . 共4.14兲 MR共N兲兲 into invariant subspaces PR共N , M , S兲共MR共N , M , S兲兲
and perform computations solely within these.
The columns of Ũ are the projections PRk “straightened The calculations were carried out with a code similar to
out” to an orthonormal set. Equation 共4.12兲 is simply the that described by Rontani et al. in Ref. 8. Table I shows
spectral decomposition of Heff. Although different in form comparisons of the present code with that of Table IV of Ref.
than most implementations in the literature 共e.g., Ref. 11兲, it 8 for various parameters using the model space
is equivalent. MR共N , M , S兲. Table I also shows the case = 1, N = 2, M
045321-8
HARMONIC OSCILLATOR EIGENFUNCTION EXPANSIONS,… PHYSICAL REVIEW B 80, 045321 共2009兲
TABLE I. Comparison of current code and Ref. 8. Figures from the latter have varying number of
significant digits. We include more digits from our own computation for reference.
= 0, and S = 0, whose exact lowest eigenvalue is E0 = 3; cf. as before, leading to artifacts for the largest values of R due
Sec. II C. We note that there are some discrepancies between to the fact that there is a finite error in the best estimates for
the results in the last digits of the results of Ref. 8. The the eigenvalues. However, in all cases there are clear, linear
spaces MR共N兲 were identical in the two approaches; i.e., the regions, in which we estimate the slope ␣. In all cases, the
number of basis functions and the number of nonzero matrix slope can be seen to decrease by at least ⌬␣ ⬇ −1 compared
elements produced are crosschecked and identical. to Fig. 3, indicating that the effective interaction indeed ac-
We have checked that the code also reproduces the results celerates the CI convergence by at least an order of magni-
of Refs. 9, 10, and 39, using the PR共N , M , S兲 spaces. Our tude. We also observe, that the relative errors are improved
code is described in detail elsewhere36 where it is also dem- by an order of magnitude or more for the lowest values of R
onstrated that it reproduces the eigenvalues of an analytically shown, indicating the gain in accuracy when using small
solvable N-particle system40 to machine precision. model spaces with the effective interaction.
Notice, that for symmetry reasons only even 共odd兲 R for
B. Experiments even 共odd兲 M yields increases in basis size
dim关PR共N , M , S兲兴, so only these values are included in the
For the remainder, we only use the energy cut spaces plots. To overcome the limitations of the two-body effective
PR共N , M , S兲. Figure 2 shows the development of the lowest interaction for higher N, an effective three-body interaction
eigenvalue E0 = E0共N , M , S兲 for N = 4, M = 0 , 1 , 2 and S = 0 as could be considered, and is hotly debated in the nuclear
function of the shell truncation parameter R, using both physics community. 共In nuclear physics, there are also more
Hamiltonians H and Heff. Apparently, the effective interac- complicated three-body effective forces that need to be
tion eigenvalues provide estimates for the ground-state ei- included.41兲 However, this will lead to a huge increase in
genvalues that are better than the bare interaction eigenval-
ues. This effect is attenuated with higher N, due to the fact 14.6
that the two-electron effective Coulomb interaction does not 14.5
take into account three- and many-body effects which be-
14.4
come substantial for higher N.
We take the Heff eigenvalues as “exact” and graph the 14.3
relative error in E0共N , M , S兲 as function of R on a logarith- 14.2
mic scale in Fig. 3, in anticipation of the relation
E0
14.1
ln共Eh − E兲 ⬇ C + ␣ ln R, ␣ = − 共k − 1 + ⑀兲. 共5.1兲
14
The graphs show straight lines for large R, while for small R 13.9
there is a transient region of nonstraight lines. For N = 5,
however, = 2 is too large a value to reach the linear regime 13.8
045321-9
SIMEN KVAAL PHYSICAL REVIEW B 80, 045321 共2009兲
Relative error
Relative error
−2
10
−3
10
−3
10
−4
10
−4
10
−5
10
−6
10 −5
10
6 8 10 12 14 16 18 20 6 8 10 12 14 16 18 20
(a) R (b) R
Relative error
−3
10
−4
10
−4
10
6 8 10 12 14 16 18 20 6 8 10 12 14 16 18 20
(c) R (d) R
FIG. 3. Plot of relative error using the bare interaction for various N, M, and S. Clear o共R␣兲 dependence in all cases.
memory consumption due to extra nonzero matrix elements. VI. DISCUSSION AND CONCLUSION
At the moment, there are no methods available that can gen-
erate the exact three-body effective interaction with suffi- We have studied approximation properties of Hermite
cient precision. functions and harmonic oscillator eigenfunctions. This in
We stress that the relative error decreases very slowly in turn allowed for a detailed convergence analysis of numeri-
general. It is a common misconception, that if a number of cal methods such as the CI method for the parabolic quantum
dot. Our main conclusion is, that for wave functions
digits of E0共N , M , S兲 is unchanged between R and R + 2, then
苸 Hk共Rn兲 falling off exponentially as 储x储 → ⬁, the shell-
these digits have converged. This is not the case, as is easily
weight function p共r兲 decays as p共r兲 = o共r−k−1兲. Applying this
seen from Fig. 3. Take for instance N = 4, M = 0 and S = 0, and
to the convergence theory of the Ritz-Galerkin method, we
= 2. For R = 14 and R = 16 we have E0 = 13.844 91 and E0 obtained estimate 共4.5兲 for the error in the eigenvalues. A
= 13.841 53, respectively, which would give a relative error complete characterization of the upper bound on the differ-
estimate of 2.4⫻ 10−4, while the correct relative error is entiability k, i.e., in 共s兲 苸 Hk, as well as a study of the con-
1.3⫻ 10−3. stant K in Eq. 共4.9兲, would complete our knowledge of the
The slopes in Fig. 3 vary greatly, showing that the eigen- convergence of the CI calculations.
functions indeed have varying global smoothness, as pre- We also demonstrated numerically that the use of a two-
dicted in Sec. III D. For 共N , M , S兲 = 共5 , 3 , 5 / 2兲, for example, body effective interaction accelerates the convergence by at
␣ ⬇ −4.2, indicating that 苸 H5共R10兲. It seems, that higher S least an order of magnitude, which shows that such a method
gives higher k, as a rule of thumb. Intuitively, this is because should be used whenever possible. On the other hand, a rig-
the Pauli principle forces the wave function to be zero at orous mathematical study of the method is yet to come.
coalesce points, thereby generating smoothness. Moreover, we have not investigated to what extent the in-
045321-10
HARMONIC OSCILLATOR EIGENFUNCTION EXPANSIONS,… PHYSICAL REVIEW B 80, 045321 共2009兲
−5 −2
10
Relative error
10
Relative error
−3
10
−4
−10
10
10
−5
10
−6
10
6 8 10 12 14 16 18 20 6 8 10 12 14 16 18 20
(a) R (b) R
Relative error
−2
10
−4
−3 10
10
−4
10
−5
10
−5
10
6 8 10 12 14 16 18 20 7 9 11 13 15 17 18
(c) R (d) R
FIG. 4. Plot of relative error using an effective interaction for various N, M, and S. Clear o共R␣兲 dependence in all cases, but notice
artifacts when R is large, due to errors in most correct eigenvalues. The R = 5 case does not contain enough data to compute the slopes with
sufficient accuracy.
crease in convergence is independent of the interaction widely studied model. Our analysis is, however, general, and
strength . This, together with a study of the accuracy of the applicable to other systems as well, e.g., quantum dots
eigenvectors, is an obvious candidate for further investiga- trapped in double wells, finite wells, and so on. Indeed, by
tion. adding a one-body potential V to the Hamiltonian H = T + U
The theory and ideas presented in this article should in we may model other geometries, as well as adding external
principle be universally applicable. In fact, Figs. 1–3 of Ref. fields.42–44
11 clearly indicate this, where the eigenvalues of 3He as
function of model space size are graphed both for bare and ACKNOWLEDGMENTS
effective interactions, showing some of the features we have
discussed. The author wishes to thank M. Hjorth-Jensen 共CMA兲 for
Other interesting future studies would be a direct com- useful discussions, suggestions, and feedback and also G. M.
parison of the direct product model space MR共N兲 and our Coclite 共University of Bari, Italy兲 for mathematical discus-
energy cut model space PR共N兲. Both techniques are com- sions. This work was financed by CMA through the Norwe-
mon, but may have different numerical characteristics. In- gian Research Council.
deed, dim关MR共N兲兴 grows much quicker than dim关PR共N兲兴,
while we are uncertain of whether the increased basis size APPENDIX: MATHEMATICAL DETAILS
yields a corresponding increased accuracy.
1. Multi-indices
We have focused on the parabolic quantum dot first be-
cause it requires relatively small matrices to be stored, due to A very handy tool for compact and unified notation, when
conservation of angular momentum, but also because it is a the dimension n of the underlying measure space Rn is a
045321-11
SIMEN KVAAL PHYSICAL REVIEW B 80, 045321 共2009兲
parameter, are multi-indices. The set In of multi-indices are 苸 In be a multi-index. If there exists a v 苸 L2共Rn兲 such that,
defined as n tuples of non-negative indices, viz, ␣ for all 苸 C⬁0 ,
= 共␣1 , . . . , ␣n兲, where ␣k ⱖ 0.
We define several useful operations on multi-indices as
follows. Let u be a formal vector of n symbols. Moreover, let
冕 Rn
共␣共兲兲共兲dn = 共− 1兲兩␣兩 冕 Rn
共兲v共兲dn , 共A9兲
共兲 = 共1 , 2 , . . . , n兲 be a function. Then, define
then ␣ ⬅ v 苸 L2 is said to be a weak derivative, or distribu-
兩␣兩 ⬅ ␣1 + ␣2 + ¯ + ␣n , 共A1兲 tional derivative, of . In this way, the weak derivative is
defined in an average sense, using integration by parts.
␣ ! ⬅ ␣1 ! ␣2 ! ¯ ␣n ! , 共A2兲 The weak derivative is unique 共up to redefinition on a set
of measure zero兲, obeys the product rule, chain rule, etc.
␣ ⫾  ⬅ 共␣1 ⫾ 1, . . . , ␣n ⫾ n兲, 共A3兲 It is easily seen that if has a classical derivative v
苸 L2共Rn兲 it coincides with the weak derivative. Moreover, if
the classical derivative is defined almost everywhere 共i.e.,
u␣ ⬅ u1␣1u2␣2, . . . ,un␣n , 共A4兲 everywhere except for a set of measure zero兲, then has a
weak derivative.
␣1 ␣1 ␣d The Sobolev space Hk共Rn兲 is defined as the subset of
␣ 共 兲 ⬅ ¯ 共兲. 共A5兲
1␣1 2␣2 d␣1 L 共Rn兲 given by
2
In Eq. 共A3兲, the result may not be a multi-index when we Hk共Rn兲 ⬅ 兵 苸 L2:␣ 苸 L2, ∀ ␣ 苸 I n, 兩␣兩 ⱕ k其.
subtract two indices, but this will not be an issue for us. 共A10兲
Notice, that Eq. 共A5兲 is a mixed partial derivative of order
兩␣兩. Moreover, we say that ␣ ⬍  if and only if ␣ j ⬍  j for all The Sobolev space is also a Hilbert space with the inner
j. We define ␣ =  similarly. We also define “basis indices” e j product
by 共e j兲 j⬘ = ␦ j,j⬘. We comment, that we will often use the no-
tation x ⬅ x and k ⬅ k to simplify notation. Thus, 共 1, 2兲 ⬅ 兺 具␣1, ␣2典, 共A11兲
␣,兩␣兩ⱕk
␣ = 共1␣1, . . . , n␣n兲, 共A6兲 and this is the main reason why one obtains a unified theory
consistent with Eq. 共A4兲. of partial differential equations using such spaces.
The space Hk共Rn兲 for n ⬎ 1 is a big space. There are some
exceptionally ill-behaved functions there; for example, there
2. Weak derivatives and Sobolev spaces are functions in Hk that are unbounded on arbitrary small
We present a quick summary of weak derivatives and re- regions but still differentiable. 共Hermite series for such func-
lated concepts needed. The material is elementary and super- tions would still converge faster than, e.g., for a function
ficial but probably unfamiliar to many readers, so we include with a jump discontinuity.兲 For our purposes, it is enough to
it here. Many terms will be left undefined; if needed, the realize that the Sobolev spaces offer exactly the notion of
reader may consult standard texts, e.g., Ref. 45. derivative we need in our analysis of the Hermite function
The space L2共Rn兲 is defined as expansions.
再
L2共Rn兲 ⬅ :Rn → C: 冕Rn
兩共兲兩2dn ⬍ + ⬁ , 冎 共A7兲 3. Proofs of propositions
We will now prove the propositions given in Sec. III and
where the Lebesgue integral is more general than the Rie- also discuss these results on mathematical terms. Recall that
mann “limit-of-small-boxes” integral. It is important that we the Hermite functions n 苸 L2共R兲, where n 苸 N0 is a non-
identify two functions and 1 differing only at a set Z negative integer, are defined by
苸 Rn of measure zero. Examples of such sets are points if
n ⱖ 1, curves if n ⱖ 2, and so on, and countable unions of n共x兲 = 共2nn ! 冑兲−1/2Hn共x兲e−x /2 ,
2
共A12兲
such. For example, the rationals constitute a set of measure
zero in R. Under this assumption, L2共Rn兲 becomes a Hilbert where Hn共x兲 is the usual Hermite polynomial.
space with the inner product A well-known method for finding the eigenfunctions of
HHO in one dimension involves writing
具 1, 2典 ⬅ 冕Rn
1共兲ⴱ2共兲dn , 共A8兲 1
HHO = a†a + ,
2
共A13兲
where the asterisk denotes complex conjugation.
where the ladder operator a is given by
The classical derivative is too limited a concept for the
abstract theory of partial differential equations, including the 1
Schrödinger equation. Let C⬁0 be the set of infinitely differ- a⬅
冑2 共x + x兲, 共A14兲
entiable functions, which are nonzero only in a ball of finite
radius. Of course, C⬁0 傺 L2共Rn兲. Let 苸 L2共Rn兲, and let ␣ with Hermitian adjoint a† given by
045321-12
HARMONIC OSCILLATOR EIGENFUNCTION EXPANSIONS,… PHYSICAL REVIEW B 80, 045321 共2009兲
兺 nk兩cn兩2 ⬍ + ⬁, 共A21兲
n=0
045321-13
SIMEN KVAAL PHYSICAL REVIEW B 80, 045321 共2009兲
2
From Lemmas 1 and 2 we extract the following important ⌽0共x兲 ⬅ −n/4e−储x储 /2 . 共A31兲
characterization of the approximating properties of Hermite
functions in d = 1 dimensions: By using Eqs. 共A4兲 and 共A18兲, we may generate all other
Proposition 1 共Approximation in one dimension兲. Hermite functions, viz.,
Let k ⱖ 0 be a given integer. Let 苸 L2共R兲 be given by
⌽␣共x兲 ⬅ ␣!−1/2共a†兲␣⌽0共x兲. 共A32兲
⬁
Given two multi-indices ␣ and , we define the polyno-
共x兲 = 兺 cnn共x兲. 共A23兲 mial P␣共兲 by
n=0
D
Then 苸 H 共R兲 if and only if 共␣ + 兲!
= 兿 P j共␣ j兲,
k
P 共 ␣ 兲 ⬅ 共A33兲
⬁
␣! j=1
045321-14
HARMONIC OSCILLATOR EIGENFUNCTION EXPANSIONS,… PHYSICAL REVIEW B 80, 045321 共2009兲
Lemma 3 共General Hermite expansions兲. In order to generate a simple and useful result for approxi-
Let 苸 L2共Rn兲, with coefficients c␣ as in Eq. 共A34兲, and mation in n dimensions, we consider the case where de-
let  苸 In be arbitrary. Assume 共a†兲 苸 L2共Rn兲. Then cays exponentially, and 苸 Hk共Rn兲, i.e., 共x兲 苸 L2共Rn兲 for
共a†兲⬘ 苸 L2共Rn兲 and a⬘ 苸 L2共Rn兲 for all ⬘ ⱕ . Moreover, all  苸 In with 兩兩 = k. In this case, we may also generalize
the following points are equivalent: Eq. 共A25兲. For this, we consider the shell weight p共r兲 defined
共1兲 共a†兲 苸 L2共Rn兲; 共2兲 for all multi-indices ␥ ⱕ , by
x 苸 L2共Rn兲; and 共3兲 兺␣␣兩c␣兩2 ⬍ +⬁.
␥ −␥
. 共A44兲
that  苸 L2共Rn兲. Let =  苸 L2. Moreover, 11 苸 L2. Proof: the only nontrivial part of the proof concerns Eq.
Since is exponentially decaying, and by the product rule, 共A42兲. Since is exponentially decaying and since 苸 Hk if
x1␥1 苸 L2 for all ␥1 ⱖ 0. By Lemma 2, x1␥111−1 苸 L2 for all and only if  苸 L2 for all , 兩兩 ⱕ k, we know that
␥1 and 0 ⬍ 1 ⱕ 1. Thus, x1␥1−e11 苸 L2. Thus, the result 兺␣␣兩c␣兩2 ⬍ +⬁ for all , 兩兩 ⱕ k. Since 兩␣兩k is a polynomial
holds as long as = e11, or equivalently = ekk for any k. of order k with terms of type a␣, a ⱖ 0, and 兩兩 = k, we
¯
To apply induction, let = x1␥111−1 苸 L2. Note that  have
苸 L2 and x¯␥ 苸 L2 for all ¯␥ = 共0 , ␥2 , . . . , ␥n兲. But by the in-
¯
duction hypothesis, x¯␥−¯ 苸 L2 for all ¯ ⱕ ¯ and all ¯␥ such 兺␣ 兩␣兩k兩c␣兩2 = 兺 a 兺 ␣兩c␣兩2 ⬍ + ⬁. 共A45兲
, 兩兩=k ␣
that ¯␥k = 0 if
¯ k = 0. This yields, using the product rule, that
x␥− 苸 L2 for all ⱕ  and all ␥ such that ␥k = 0 if k On the other hand, since a ⱖ 0 and the sum over  has
= 0, which was the hypothesis for n dimensions, and the finitely many terms, 兺␣兩␣兩k兩c␣兩 ⬍ +⬁ implies 兺␣␣兩c␣兩2
proof is complete. Notice, that we have proved that 共a†兲 ⬍ +⬁ for all , 兩兩 = k, and thus 苸 Hk since was exponen-
苸 L2 as a by-product. 䊏 tially decaying. 䊏
045321-15
SIMEN KVAAL PHYSICAL REVIEW B 80, 045321 共2009兲
8
M. Rontani, C. Cavazzoni, D. Bellucci, and G. Goldoni, J. 1972兲.
Chem. Phys. 124, 124102 共2006兲. 28 M. Hoffmann-Ostenhof, T. Hoffmann-Ostenhof, and H. Strem-
9
A. Wensauer, M. Korkusinski, and P. Hawrylak, Solid State nitzer, Commun. Math. Phys. 163, 185 共1994兲.
Commun. 130, 115 共2004兲. 29
H. Yserentant, Numer. Math. 98, 731 共2004兲.
10 S. A. Mikhailov, Phys. Rev. B 66, 153313 共2002兲.
30 I. Babuska and J. Osborn, Handbook of Numerical Analysis:
11
P. Navrátil, J. P. Vary, and B. R. Barrett, Phys. Rev. C 62,
Finite Element Methods 共North-Holland, London, 1996兲, Vol. 2.
054311 共2000兲. 31 I. Babuska and J. Osborn, SIAM 共Soc. Ind. Appl. Math.兲 J. Nu-
12 R. Ram-Mohan, Finite Element and Boundary Element Applica-
mer. Anal. 24, 1249 共1987兲.
tions in Quantum Mechanics 共Oxford University Press, New 32 M. B. Tavernier, E. Anisimovas, and F. M. Peeters, Phys. Rev. B
York, 2002兲.
13 A. Kumar, S. E. Laux, and F. Stern, Phys. Rev. B 42, 5166 74, 125305 共2006兲.
33 I. Babuska and J. Osborn, Math. Comput. 52, 275 共1989兲.
共1990兲. 34 D. Klein, J. Chem. Phys. 61, 786 共1974兲.
14
M. Macucci, K. Hess, and G. J. Iafrate, Phys. Rev. B 55, R4879
35 S. Kvaal, Phys. Rev. C 78, 044330 共2008兲.
共1997兲.
15 P. A. Maksym and N. A. Bruce, Physica E 1, 211 共1998兲. 36
S. Kvaal, arXiv:0810.2644 共unpublished兲.
16
T. Ezaki, N. Mori, and C. Hamaguchi, Phys. Rev. B 56, 6428
37
J. S. Ball, SIAM 共Soc. Ind. Appl. Math.兲 J. Numer. Anal. 40,
共1997兲. 2311 共2002兲.
17
P. Maksym, Physica B 249-251, 233 共1998兲. 38
T. Schucan and H. Weidenmüller, Ann. Phys. 76, 483 共1973兲.
18 39 S. A. Mikhailov, Phys. Rev. B 65, 115312 共2002兲.
M. B. Tavernier, E. Anisimovas, F. M. Peeters, B. Szafran, J.
Adamowski, and S. Bednarek, Phys. Rev. B 68, 205305 共2003兲. 40 N. F. Johnson and M. C. Payne, Phys. Rev. Lett. 67, 1157
19 M. Taut, Phys. Rev. A 48, 3561 共1993兲.
共1991兲.
20 M. Taut, J. Phys. A 27, 1045 共1994兲. 41 E. Caurier, G. Martinez-Pinedo, F. Nowacki, A. Poves, and A. P.
21 D. Pfannkuche, V. Gudmundsson, and P. A. Maksym, Phys. Rev.
Zuker, Rev. Mod. Phys. 77, 427 共2005兲.
B 47, 2244 共1993兲. 42 M. Helle, A. Harju, and R. M. Nieminen, Phys. Rev. B 72,
22 M. Hoffmann-Ostenhof, T. Hoffmann-Ostenhof, and H. Strem-
205329 共2005兲.
nitzer, Phys. Rev. Lett. 68, 3857 共1992兲. 43 M. Braskén, M. Lindberg, D. Sundholm, and J. Olsen, Phys.
23 T. Kato, Commun. Pure Appl. Math. 10, 151 共1957兲.
Rev. B 61, 7652 共2000兲.
24 R. J. Bartlett and M. Musiał, Rev. Mod. Phys. 79, 291 共2007兲. 44 S. Corni, M. Braskén, M. Lindberg, J. Olsen, and D. Sundholm,
25 J. Boyd, J. Comput. Phys. 54, 382 共1984兲.
Phys. Rev. B 67, 045313 共2003兲.
26 E. Hille, Duke Math. J. 5, 875 共1939兲. 45 L. Evans, Partial Differential Equations, Graduate Studies in
27
S. Raimes, Many-Electron Theory 共North-Holland, London, Mathematics 共AMS, Providence, RI, 1998兲, Vol. 19.
045321-16