Research Collection Doctoral Thesis First-Principles Simulations of Multi-Orbital Systems with Strong Electronic Correlations Author(s): Surer, Brigitte Publication Date: 2011 Permanent Link: https://doi.org/10.3929/ethz-a-6665042 Rights / License: In Copyright - Non-Commercial Use Permitted This page was generated automatically upon download from the ETH Zurich Research Collection . For more information please consult the Terms of use . ETH Library
153
Embed
Strong Electronic Correlations First-Principles ... · 2.2 Models of strongly correlated systems . . . . . . . . . . . . . . 10 ... of the electronic structure making the development
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Research Collection
Doctoral Thesis
First-Principles Simulations of Multi-Orbital Systems withStrong Electronic Correlations
Figure 4.2: Left: One peak and double peak structures at the frequency ω0.
The peaks in the double peak structure are separated by a gap of width ∆ω.
Depending on ω0 and ∆ω the differences ∆G in the Green’s function may be
too small to resolve the double peak structure. Right: Difference in the Green’s
function between a single peak and double peak-structure with a gap of ∆ω = 1 eV
for β = 20 eV−1. The peaks are located at ω0. Given statistical errors of the order
10−4 (line) there are only a few points carrying the information about the difference
between the double peak and single peak in the spectral function.
For simplicity the difference ∆A(ω) shall be a rectangle of width ∆ω at posi-
tion ω0 and we do not care about the normalization of A(ω). We can now ask
how the corresponding Green’s function will differ. Since the transformation
54
Analytical continuation
is linear the difference ∆G is:
∆G(τ) =
∫ ∞
−∞
dωK(τ, ω)∆A(ω). (4.3)
For a rectangle of width ∆ω = 1 eV we compute ∆G(τ) at an inverse temper-
ature β = 20 eV−1 and present ∆G for ω0 = 0.5, 1.5, 2.5, 3.5, 4.5 eV in Figure
4.2.
It can be seen that ∆G becomes small for values away from τ = 0, the
larger ω0 the steeper the decay of ∆G. Since there is a certain statistical
error associated with the values of G there will be only a few values which
can have the information about the double peak structure encoded. In our
simulations this statistical error is typically of the order 10−5 − 10−4 which
we have indicated in Figure 4.2 by a threshold line. If one is particularly
interested in reproducing a particular feature in a PES-spectra - especially
those which had not been reproduced by any previous QMC-simulation - it
is advisable to estimate the required data accuracy for a desired number of
information-relevant Green’s function values in the proposed way. In combi-
nation with test-runs one can extrapolate the required computing resources
using the fact that a ten times smaller error bar requires 100 times more
measurements according to Eq.(3.3). Having made these preliminary consid-
erations we will now outline three methods how the spectral function can be
computed from the imaginary-time Green’s function. We have used in this
thesis three approaches - the Maximum Entropy Method (MaxEnt) [56, 57]
the Stochastic optimization method [58] and the analytic continuation with
Pade approximants [59].
55
4.2 Maximum Entropy method
4.2 Maximum Entropy method
The Monte Carlo simulation yields an estimate G(τ) of the exact Green’s
function G(τ). Given a spectral function A(ω) we can ask about the prob-
ability P[A, G
]that A generates the Green’s function G(τ) for which we
provide the Monte-Carlo estimate G(τ). Bayes theorem [60] connects this
posterior probability to the prior probability P [A] and likelihood function
P[G|A
]
P[A, G
]∝ P
[G|A
]P [A] (4.4)
For the likelihood function P[G|A
]∼ e−L we use a least-square fitting mea-
sure
L =1
2χ2 =
1
2
∑
k
Gk − Gk
σ2k
(4.5)
in order to quantify how well G(τ) with statistical errors σk is reproduced by
G(τ) which is the Green’s function obtained by back-transforming the given
spectral function A(ω). Maximizing the likelihood function is equivalent to
minimizing the deviation measure L. In principle we could just stop here,
assume P[A, G
]∝ P
[G|A
]and simply select the spectral function A(ω)
which maximizes the likelihood function. We will do so for the Stochastic
optimization method of spectral analysis discussed in Section 4.3. Choosing
the spectral function, which produces a Green’s function G(τ) fitting G(τ)
best will still be different from the exact value G(τ). As we have seen from
our considerations in Section 4.1 tiny changes in G(τ) can result in large
differences in A(ω). Therefore, we will have most likely artificial features in
A(ω) which are not present in the ”true” spectral function. In the Maximum
Entropy method (MaxEnt) we seek to avoid the overfitting of the data via
regularization. In MaxEnt the prior probability is chosen as P [A] ∼ eαS
with the entropy S
S = −∫
dωA(ω) ln
[A(ω)
m(ω)
]
(4.6)
and the default model m(ω) serving as a reference system. By choosing
the entropy as a regularization function we ensure that correlations are not
introduced via the regularizer. Correlations are only produced if the data
supports them.
With the above choices for the likelihood function and prior probability we
can now explicitly maximize the posterior probability P[A, G
]by maximiz-
56
Analytical continuation
ing:
Q = αS − L. (4.7)
The parameter α plays the role of a regularization parameter and controls the
competition between the entropic solution and the data constraints imposed
by the fitting.
We will now outline the MaxEnt algorithm following [57]. As an input we
provide the Monte-Carlo Green’s function Gj = G(τj), j = 1, ..., Nd with
statistical errors σj . We decide on the frequency width ∆ωi and number of
frequency points N for which we want to obtain the spectral density Ai =
A(ωi)∆ωi. The Kernel in discretized form is given by the matrix:
Tij =e−τjωi
1 + e−βωi(4.8)
and yields the ”mock” Green’s function
Gj =∑
i
TijAi (4.9)
How well G matches the measured data is specified by the least-square fitting
Eq. (4.5). As a default model we use mi = m(ωi)∆ωi = const. and use the
entropy:
S =∑
i
Ai −mi − Ai log(Aimi
)
. (4.10)
We decompose the Kernel T using a singular value decomposition
T = V ΣUT (4.11)
Σ is a diagonal matrix diag(γ1, ...., γs, 0, ..., 0) with s non-zero singular values
γi. We can restrict V , Σ and U to this s-dimensional subspace:
T = Vs︸︷︷︸
Nd×s
Σs︸︷︷︸
s×s
UTs
︸︷︷︸
s×N
(4.12)
which makes the calculation of maximizing Q much more efficient.
In order to maximize Q we compute the derivative with respect to A:
α∇S −∇L = 0 (4.13)
Bryan showed that this equation can be solved using a Newton method pro-
ceeding in steps δu defined via [57]:
(αI +MK) δu = −αu− g (4.14)
57
4.2 Maximum Entropy method
with
M = ΣsVTs diag
(1
σ1, ...,
1
σNd
)
VsΣs
K = UTs diag (A1, ..., AN)Us
u = UTv, vi = lnAimi
g = ΣV Tw, wi =∆Gi
σ2i
(4.15)
At step p we update up+1 = up + δup and stop if δuTKδu = 0. We must
take care that the step length does not become too large. If δuTKδu > 1 we
solve instead of Eq. (4.14).
((α + µ)I +MK) δu = −αu− g (4.16)
choosing µ large enough to ensure δuTKδu . 1.
The final u yields the spectral function
A = meUTs u (4.17)
for a given value of α.
The optimal choice of α can again be found using Bayes Theorem. The
conditional probability P[α|G,m
]is proportional to
P[α|G,m
]∝ Πi
(α
α + λi
)1/2
eQP [α] (4.18)
with λi the eigenvalues of MK. The optimal value for α is found for:
−2αoptS =∑
i
λiλi + αopt
. (4.19)
Above equation can be viewed as a count Ng of good measurements, since
each λi ≫ α increases the left hand side by one.
Instead of using only this single αopt parameter we average over a range of
reasonably probable α according to their probability Pα = Πi
(α
α+λi
)1/2
eQ:
〈A〉 =∑
α PαAα∑
α Pα. (4.20)
For this we choose a step length ∆α and consider those αt for which lnPαt<
lnPαopt+∆ for a chosen threshold ∆.
58
Analytical continuation
4.3 Stochastic optimization method of spectral analysis
The stochastic optimization method of spectral analysis has been developed
by Mishchenko et al. [58]. and is based on χ2-minimization only, i.e., it does
not introduce any regularization. Its advantages is that unlike MaxEnt it
can restore sharp features in the spectral function and does not rely on a
discretized frequency space.
We define an optimization procedure in which we update the proposed spec-
tral function stochastically. We generate several valid spectral functions
which fulfill the condition that the back-transformed Green’s function de-
viates from the measured Green’s function at most by a threshold Dmin
We consider the deviation function
D(A) =
∣∣∣G(τ)− G(τ)
∣∣∣
G(τ)(4.21)
where we ensure by dividing by G(τ) that the Green’s function is also closely
approximated where it has very small values, e.g. for an insulating system
at intermediate τ .
We represent the proposed spectral function A(ω) as composed of rectangular
contributions:
˜A(ω) =K∑
t=1
χRt(ω) (4.22)
with rectangles Rt = ht, wt, ct defined as
χRt(ω) =
ht, if ω ∈ [ct − wt/2, ct + wt/2]
0, otherwise(4.23)
with positive values for the height ht and width wt and centers ct within the
considered frequency range.
The normalization is fulfilled by imposing:
K∑
t=1
htwt = 1 (4.24)
for a given configuration C = R1, ..., RK of K rectangles.
For the spectral function A(ω) of a given configuration C the corresponding
Green’s function is computed according to Eq. (4.9). For each configura-
tion we are therefore able to measure the deviation of the back-transformed
Green’s function to the measured Green’s function using Eq. (4.21).
59
4.3 Stochastic optimization method of spectral analysis
We will now discuss how we update the configuration of rectangles and ac-
cording to which criterion we accept new configurations within our optimiza-
tion procedure. There are updates changing only the shape of the rectangles
but keeping the total number fixed and updates for changing the number of
rectangles of a configuration. The moves consist of
1. Shifting a rectangle.
2. Reshaping a rectangle by keeping its area fixed, i.e. changing the width
and height simultaneously.
3. Reshaping two rectangles, such that the sum of both areas stay fixed.
4. Adding a new rectangle and simultaneously reducing the weight of an-
other randomly chosen rectangle accordingly.
5. Removing a rectangle by simultaneously enlarging another randomly
chosen rectangle accordingly.
6. Splitting a rectangle and shifting the two parts.
7. Glueing two rectangles together and reducing the width to again fulfill
the normalization condition.
We update a new configuration r′ with probability:
Pr→r′ =
1 if D [C(r′)] < D [C(r)]f (D [C(r)] /D [C(r′)]) , otherwise
(4.25)
with f(x) = x1+d using d ∈ [1, 2]. We therefore also allow for updates which
increase the deviation D, because we want to avoid getting stuck in local
minima and want to ensure that we can in principle generate any positive
and bounded spectral function A(ω). The stochastic update procedure is
very similar to the updates in Monte-Carlo algorithms. However, since we
are dealing with an optimization procedure, there is no need to fulfill the de-
tailed balance condition, which would be required if we measured observables
according to a given statistical ensemble.
60
Analytical continuation
4.4 Analytic continuation with Pade approximants
Instead of an optimization procedure such as MaxEnt or the Stochastic op-
timization method we can also interpolate the data with a polynomial and
analytically continue the polynomial. We present here the analytic continua-
tion with Pade approximants which are based on an interpolation of rational
polynomials [59]. By interpolating our data we do not take into account
any statistical errors which may lead to artificial features in the spectral
function. Nevertheless, the procedure turns out to be quite reliable for fre-
quencies close to the Matsubara axis, where the quality of the QMC data is
sufficiently good. We consider a rational polynomial r(x) of the form
r(x) = f0 +a1(x− x0)
1 +a2(x− x1)
1 +a3(x− x2)
1 +a4(x− x3)1 + ...
(4.26)
which interpolates given values of a function f(x). We choose N points of
the function of f(x) to be interpolated at the points S = x1, x2, ..., xN. Ina first step we define the initial reduced function g(0)(x) given the set S:
g(0)(xi) = f(xi)− f(x0), xi ∈ S. (4.27)
We will recursively construct the coefficients a1, a2, ..., aN specifying the
rational polynomial. At step j we compute
aj =g(j−1)(xj)
(xj − xj−1)(4.28)
and set
g(j)(xi) = ajxi − xj−1
g(j−1)(xi)− 1, i 6= j − 1 (4.29)
.
We repeat this procedure until for all xi ∈ S we find that either g(j−1)(xi) =
0 or =∞. From the coefficients a1, a2, ..., aN we construct recursively the
polynomial r(x) for any desired values x.
In order to evaluate the polynomial for a specific value of x we compute the
continued fraction in the following iterative way: We start with q1(x)/q0(x) =
1, i.e., q−1(x) = 0, q0(x) = 1 and compute
qi+1(x) = qi(x) + ai+1(x− xi)qi−1(x), i = 1, ..., t = N − 1 (4.30)
61
4.5 Benchmarks
or equivalently:
qi+1(x)/qi(x) =1 + ai+1(x− xi)qi(x)/qi−1(x)
(4.31)
From the final continued fraction qt+1(x)/qt(x) we compute the interpolating
rational polynomial r(x) as r(x) = qt+1(x)/qt(x)− 1 + f0.
We apply this interpolation scheme to the Green’s function G(iω). By con-
struction of the interpolating polynomial we will obtain the measured values
G(iωn) on the Matsubara frequency points
iπβ, i3π
β, ..., i (2N+1)π
β
. We are
now able to analytically continue G(iωn) by evaluating r(x) at x = iω + iδ
which yields an estimation of G(ω). We can in principle provide any de-
sired frequency range and resolution. However, the further away we are from
the interpolation points iωn the less trustworthy the generated values will
be. From the analytically continued Green’s function G(ω) we compute the
spectral function
A(ω) = −1
πImG(ω). (4.32)
Since we do not have a measure of the error for the values away from the
interpolation points, we have to examine whether the result is reasonable.
As a first check we study how including a different number N of Matsubara
frequency points will affect the result. Features which exist only for one
particular choice of N can in this way be detected as artificial. As a second
check we compare the result to MaxEnt and the stochastic optimization
method result.
4.5 Benchmarks
We found Pade to work well to resolve quasi-particle peaks for Kondo impu-
rity problems. It agrees nicely with MaxEnt and the stochastic optimization
method as shown in Figure 4.3 in the energy range [−1, 1] eV. If we compare
the stochastic optimization method to MaxEnt the spectral function looks
very noisy and it is not clear to decide which features are physical and which
artificial.
The three methods are very different in how they favor regularization ver-
sus fulfilling the data constraints, i.e. in how explicitly the data is taken.
We can arrange the methods along a line between these two extremes as
depicted in Figure 4.4. Pade means directly inverting the Green’s function,
the Stochastic optimization method is based on likelihood-minimization only
62
Analytical continuation
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
−10 −8 −6 −4 −2 0 2 4 6 8 10
DO
S
Energy (eV)
PadeStochastic
MaxEnt
0
1
−1 0 1
Figure 4.3: Spectral function of Co on Cu, µ = 29 eV obtained via different
analytic continuation methods.
and MaxEnt balances the minimizing of the likelihood and maximizing of the
entropy with the parameter α. If a feature in the spectral function persists
over a wide range on this axis, it is very likely that the feature is really sup-
ported by the data. In other words it will reflect that the data statistics is
good enough such that the information of this feature is encoded in enough
data points as outlined in our considerations in Section 4.1. In conclusion we
suggest to examine first what the accuracy requirements for a studied energy
range are and to make out via a comparison of different methods whether
the data accuracy is good enough.
MaxEnt
large α small α
stochastic optimization
method
Pade
regularization data constraints
Figure 4.4: Characterization of the different analytic continuation methods in
terms of the two extremes of regularization and fulfillment of the data constraints.
63
4.5 Benchmarks
64
Chapter 5
Transition metal impurities on
metal surfaces
5.1 Introduction
Quantum impurity models are the most basic models capturing the compe-
tition between localization and delocalization and are thus very instructive
to learn about the interplay of atomic and itinerant physics in strongly cor-
related systems. With the quantum impurity solver used in this thesis, local
degrees of freedom can be incorporated, which allows to study the phenomena
originating from orbital degrees of freedom. Most importantly, the quantum
impurity solver used provides a numerically exact solution to quantum im-
purity models. It is therefore possible to directly test the accuracy of the
first-principles calculations providing the material specific input parameters.
The approach to incorporate material specific effects to Model-Hamiltonians
describing strongly correlated systems is also used in lattice calculations,
e.g. realized in “LDA+DMFT” first-principles simulations. As will be ex-
plained in Section 6 these simulations involve the DMFA and for a given
system it is often not investigated what the effect of this approximation is
and whether deviations from the experimental data should be accounted to
this approximation or other effects. By taking one step back to investigations
of quantum impurity systems we can benchmark the predictive power of the
“DFT+Model-Hamiltonian” approach.
65
5.2 Multiplet features in Fe on Na
5.2 Multiplet features in Fe on Na
5.2.1 Introduction
To the presented study the following persons have contributed: B. Surer,
T. Wehling, A. Belozerov, A. Poteryaev, A. Lichtenstein, M. Troyer, A.
Lauchli and P. Werner. Our aim in this study is to use the ”DFT+Model-
Hamiltonian” approach in combination with the Krylov implementation of
the Hybridization solver to reproduce the PES-spectra of a system, which
exhibits very rich features in the spectra. Carbone et al. [61] studied Fe
atoms on different alkali metals. The hybridization varies with the metallic
host - the heavier the alkali ions the lower the electron density and hence
also the weaker the hybridization. In Figure 5.1 a sketch is shown how the
spectral density varies with the strength of hybridization. At the bottom the
E
DOS
Figure 5.1: Sketch of the densities of states of a Fe impurity on different alkali
hosts. As the hybridization varies along the alkali series the DOS changes from
atomic-like (bottom) to a system with quasi-particle peak and Hubbard bands
(top).
hybridization is very weak (host: Cs) and the atomic-like features are hardly
changed. In the intermediate hybridization regime (host: Na) a quasi-particle
peak develops and some of the peaks get broadened to form an incoherent
feature, but still there are features reminiscent of the atomic system. At
strong hybridization all the atomic features are dissolved in a broad Hub-
bard band and the intense scattering processes between the impurity and
the conduction electrons is reflected in the distinct quasi-particle peak (host:
Li).
66
Transition metal impurities on metal surfaces
5.2.2 Model and Computational method
We choose to study the intermediate case represented by Fe on Na having a
feature rich spectra, which exhibits multiplet features as well effects from the
coupling to the bath. Since schemes in the density-density approximation fail
to correctly reproduce multiplet features we study the multi-orbital quantum
impurity system using the full Coulomb interaction matrix as defined in Eq.
(2.7). We study the system for the Coulomb interaction parameters U = 4 eV
and J = 0.9 eV and apply the Krylov hybridization solver which is able to
treat the local Hamiltonian on the five orbitals exactly. We compare our
results to ED-calculations of three cases. First, we address the local problem
only and identify which transitions the peaks in the spectra stand for. Second,
we couple the 5-orbitals to one bath sites each (’5+5’) as well two bath
sites each (’5+10’). This allows us to investigate the effect of coupling the
impurity with more bath sites, in which case the CTQMC-result represents
the coupling to an infinite number of bath sites. For the total occupation of
the Fe atom on a Na host DFT yields n . 7 and we therefore consider total
densities n ∼ 6− 7.
5.2.3 Results
5.2.3.1 Identification of the transitions
The spectral function has resonances at the excitation energies of electron
removal and addition processes as shown in Section 2.5. The energy range
below the Fermi-level (EF = 0) corresponds to the transitions dn → dn−1,
the energy range above EF to transitions dn → dn+1. The initial state is
the ground-state whose spin S and angular momentum L are in line with
Hund’s rule. In a multi-orbital system there are different possibilities of how
the electrons are arranged in the final dn±1-configuration depending on from
which (to which) orbital we remove (add) the electron. The final states can be
classified according to their angular momentum L′ and spin S ′. For the given
symmetry of the local Hamiltonian Hloc the set of states of a given (L, S)
value will be degenerate and therefore the peak in the spectra will represent
transitions to not only a single state but to a multiplet. In Figure 5.2 we show
the spectra of the isolated Fe atom with an occupation of n = 7. Since it is
not coupled to a bath, the system is solely described by Hloc. In the atomic
problem we can associate quantum numbers n, l,m to the eigenenergies Enlm.
67
5.2 Multiplet features in Fe on Na
0
0.5
1
1.5
2
−7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5
DO
S
E (eV)
(3,3/2) −> (3,1)
(3,3/2) −> (4,1)
(3,3/2) −> (1,1)
(3,3/2) −> (5,1)
(3,3/2) −> (2,1)
(3,3/2) −> (2,2)(3,3/2) −> (3,1)
(3,3/2) −> (1,1)
(L,S) −> (L’,S’)
Figure 5.2: Spectra of the isolated Fe atom: The peaks mark the transition from
the (L,S) = (3, 3/2) ground state to the final states (L′, S′). The specific (L′, S′)
multiplet is indicated at the corresponding peak.
We subtract the ground-state energy EGS from the eigenenergies to obtain
all energy differences ∆E = En′l′m′ − EGS. We determine the transitions
by identifying the peak position Ep with the excitation energies ∆E which
allows us to tag the angular momentum L and spin S to the final state. In
the ground-state the seven electrons arrange to a spin S = 3/2 and angular
momentum L = 3. The peak closest to EF represents the transition to
S ′ = 2 and L′ = 2 state, which is the ground-state of a d6-configuration.
Other final d6 configurations have a higher excitation energy and therefore
lie at higher excitation energies further away from EF . We have 5 different
final multiplets with S = 1 and L = 1, 2, 3, 4, 5. The transitions to the d8
configurations have large excitation energies due to the Coulomb repulsion
and because the charge cannot be screened. In the presence of itinerant
electrons we expect that the environment will react to an additional charge.
As a next step we will therefore study what happens if we place the system
in a metallic host.
68
Transition metal impurities on metal surfaces
5.2.3.2 Effect of the bath
The coupling between the Fe atom and the Na host is encoded in the hy-
bridization function ∆. To obtain ∆ we performed DFT calculations using
a generalized gradient approximation (GGA) [62] as implemented in the Vi-
enna Ab-Initio Simulation Package (VASP) [63] with projector augmented
waves basis sets (PAW) [64,65]. For the simulation of the Fe impurity on the
(100) bcc Na surface we used a 2 × 2 supercell x 7 layer slab as a supercell
structure. The Fe adatom and the three topmost Na layers were relaxed until
the forces acting on each atom were below 0.02 eVA−1. Using the projection
method as outlined in Section 2.4.1 we extracted the hybridization functions
from our DFT calculations. In Figure 5.3 we show Im∆(ω) for Fe on Na.
The hybridization is especially strong at large positive frequencies. In the
ED calculation the parameters Vp and bath site energies ǫp are chosen such
to approximately cover the hybridization function. In Figure 5.3 we show the
−0.16
−0.14
−0.12
−0.1
−0.08
−0.06
−0.04
−0.02
0
−5 −4 −3 −2 −1 0 1 2
Im ∆
(eV
)
E (eV)
dxydx
2−y
2
dxz,dyzdz
2
0
0.5
1
1.5
2
2.5
3
−7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5
DO
S
E (eV)
isolated Fe atom5 bath sites10 bath sitescontinuum
Figure 5.3: Left: Hybridization function Im∆(ω) for Fe on Na. Right: Evolution
of the spectral function of Fe on Na with increasing number of bath sites.
spectra obtained by ’5+5’ and ’5+10’ ED-calculations , as well the CTQMC
result (’5+∞’). At negative frequencies the peaks of the ED-calculation get
shifted by a constant shift of around 1 eV compared to the atomic case. This
implies that in the Green’s function G(iωn) = (iωn + ǫd −∆− Σ)−1 the term
−∆ − Σ is well represented by a static value and the level ǫd gets shifted to
the position ǫd −∆0 − Σ0. Apart from the shift the peak structure remains
qualitatively the same in the ED-calculation. At high frequencies the spec-
trum changes substantially. The strong hybridization allows the conduction
electrons to efficiently screen the additional charge. Including more bath sites
69
5.2 Multiplet features in Fe on Na
(’5+10’) does not change the spectra much and also the analytically contin-
ued CTQMC data (’5 +∞’) is in very good agreement with the ED-data.
The transition to the L′ = 2, S ′ = 2 state is clearly visible in the CTQMC
spectral function, whereas the different S = 1 multiplets get combined into
one broad feature. It cannot be made out whether this is due to the fact
that the bath strongly mixes the impurity states leading to a broadening or
whether this is due to the smoothing inherent to MaxEnt.
5.2.3.3 Comparison to PES
0
0.5
1
1.5
2
2.5
3
3.5
−7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5
DO
S
E (eV)
CTQMCED 5+5
ED 5+10
0
0.5
1
1.5
2
2.5
3
3.5
−7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5
DO
S
E (eV)
CTQMCED 5+5
ED 5+10
Figure 5.4: Spectral function of Fe on Na for the chemical potential µ = 22 eV
(left) and µ = 24 eV (right). The data for µ = 23 eV is shown in Figure 5.3.
In our simulations we have to adapt the chemical potential for a given total
density for which we use the DFT estimate n . 7. We performed calculations
for µ = 22, 23, 24 eV which yields total densities n = 6, 6.9, 7.0. For all these
choices we find very good agreement between ED and CTQMC see Figure 5.4.
We find best agreement with the the experimental Photo-emission spectrum
of Fe on Na for the density n = 6.9, i.e. µ = 23 eV. The comparison is shown
in Figure 5.5. In both experimental and computed spectra we find a quasi-
particle peak at EF , a multiplet peak at −0.5 eV - which is the transition to
the L′ = 2, S ′ = 2 multiplet - and a broad Hubbard band originating from
many multiplets at −2 to −4 eV. The spectra obtained from ED can even
resolve the double peak structure in the Hubbard band.
70
Transition metal impurities on metal surfaces
0
0.5
1
1.5
2
2.5
3
3.5
4
−4 −3 −2 −1 0
DO
S
E (eV)
H M QP
CTQMCED 5+10
PES
Figure 5.5: Comparison of the computed spectral function with the experimental
PES data [61]. Quasi-particle peak (QP), Multiplet (M) at −0.5 eV and Hubbard
band (H) between −2 and −4 eV can be reproduced.
5.2.4 Conclusion
We conclude that applying the material specific “DFT+Model Hamiltonian”
approach in the full Coulomb formulation applied to the Fe impurity on Na
system allows to reasonably reproduce the PES-spectra. Since the spectra
is sensitive to the choice of µ, which we cannot determine ab-initio with the
desired precision the simulations were not fully ab-initio and required us to
compare a certain density range to the experimental data. Improvements to
better predict the chemical potential are needed. We have studied the effect
of the reservoir of conduction electrons in the Na host and shown that it acts
to screen additional charges on the Fe atom. Concerning the methods we find
the combination of ED and CTQMC very powerful, because it enables us to
consider a continuous bath and to still identify the underlying transitions in
the spectra, as well to judge the quality of the analytical continuation of the
CTQMC data.
71
5.3 Kondo physics in Co on Cu and Co in Cu
5.3 Kondo physics in Co on Cu and Co in Cu
B. Surer, T. Wehling, A. Wilhelm,
A. Lichtenstein, M. Troyer, A. Lauchli and P. Werner
submitted to Phys. Rev. B
5.3.1 Introduction
We investigate the electronic structure of cobalt atoms on a copper surface
and in a copper host by combining DFT calculations with a numerically
exact CTQMC treatment of the five-orbital impurity problem. In both cases
we find low energy resonances in the DOS of all five Co d-orbitals. The
corresponding self-energies indicate the formation of a Fermi liquid state at
low temperatures. Our calculations yield the characteristic energy scale – the
Kondo temperature – for both systems in good agreement with experiments.
We quantify the charge fluctuations in both geometries and suggest that
Co in Cu must be described by an Anderson impurity model rather than
by a model assuming frozen impurity valency at low energies. We show
that fluctuations of the orbital degrees of freedom are crucial for explaining
the Kondo temperatures obtained in our calculations and and measured in
experiments.
Figure 5.6: Sketch of Co impurities in a Cu host. We consider the two cases: (i)
the impurity buried in the bulk (Co in Cu), and (ii) on top of the Cu layer (Co on
Cu).
The Kondo effect, arising when localized spins interact with a metallic envi-
ronment, is a classic many body problem [66]. At present, idealized models
dealing with, for instance, a single spin degree of freedom screened by a sea
of conduction electrons are well understood. Spin S = 1/2 Kondo models or
single orbital Anderson impurity models have been widely considered to de-
scribe magnetic impurities with open d-shells in metallic environments and
72
Transition metal impurities on metal surfaces
proved helpful in qualitative discussions [67–74]. As realized, however, by
Nozieres and Blandin in 1980 [75], such idealized models may ignore impor-
tant aspects of the nature of transition metal impurities as they disregard
orbital degrees of freedom. This makes comparisons between theory and ex-
periment often very difficult. More realistic models accounting for the orbital
structure, Hund’s rule coupling, non-spherical crystal fields and an energy-
and orbital-dependent hybridization of the impurity electrons with the sur-
rounding metal are theoretically very demanding due to the multiple degrees
of freedom and multiple energy scales involved.
In recent years different attempts have been made to address this problem.
For the classic example of Fe in Au, which has been experimentally studied
since the 1930s, a model describing the low energy physics has been derived
[76] by comparing numerical renormalization group (NRG) calculations to
electron transport experiments. The Kondo temperature TK , below which
the impurity spin becomes screened and a Fermi liquid develops, served as
a fitting parameter in this study. A scaling analysis of multiple Hund’s
coupled spins in a metallic environment showed that Hund’s rule coupling
can strongly quench the formation of Kondo singlet states [77]. For highly
symmetric systems like Co adatoms on graphene [78] or Co-benzene sandwich
molecules in contact to metallic leads [79], the orbital degree of freedom has
been suggested to control Kondo physics down to the lowest energy scale.
However, a general strategy to assess which degrees of freedom are involved
in the formation of low energy Fermi liquids around magnetic impurities in
metals is still lacking.
Co atoms coupled to Cu hosts present another experimentally extensively
studied system which has been interpreted in terms of Kondo physics [68,
69, 71–74, 80]. Theoretical descriptions of this system have often been based
on single orbital Anderson impurity models [67–70] or Kondo models [71]
and the role of orbital fluctuations in these systems has remained rather un-
clear. Recently developed CTQMC [46] approaches allow to describe the full
orbital structure of magnetic impurities in metallic hosts, while accounting
for all electron correlations in a numerically exact way. So far, however,
such CTQMC studies have been limited to rather high temperatures [81],
well above typical Kondo scales on the order of 10K to 500K. The realistic
description of transition metal Kondo systems thus remains a long standing
open problem in solid state physics.
Here, we employ the recently developed Krylov CTQMC method [47] in com-
73
5.3 Kondo physics in Co on Cu and Co in Cu
bination with DFT based first-principles calculations to achieve an ab-initio
description of two archetypical Kondo systems: Co adatoms on a Cu (111)
surface, as well as Co impurities in bulk Cu (see Figure 5.6). We consider the
energy dependent hybridization of the impurities with the surrounding host
material as well as the full local Coulomb interaction and find low energy
resonances developing in the spectral function as the temperature is lowered.
Such resonances are found in all impurity 3d-orbitals and our calculations in-
dicate that spin- and orbital-fluctuations are crucial for the formation of low
energy Fermi liquids involving all impurity 3d-orbitals. We also demonstrate
the intermediate-valence character of the Co impurity in bulk, which implies
that the physics cannot be correctly described by a low-energy Kondo model
which neglects charge fluctuations.
5.3.2 Model and Computational method
A realistic description of the Co atoms on Cu (111) and in bulk Cu including
all five Co 3d orbitals can be formulated in terms of a multi-orbital Ander-
son impurity model Eq.(2.9) with the full four-fermion Coulomb interaction
Eq.(2.7).
To specify the parameters of impurity models describing Co on Cu (111) as
well as Co in bulk Cu we performed first-principles calculations.
DFT calculations were performed to obtain relaxed geometries and the hy-
bridization functions for single Co atoms in Cu and on a Cu (111) surface.
The DFT calculations have been carried out using a generalized gradient
approximation (GGA) [62] as implemented in the Vienna Ab-Initio Sim-
ulation Package (VASP) [63] with projector augmented waves basis sets
(PAW) [64, 65]. For the simulation of a cobalt impurity in bulk Cu we em-
ployed a CoCu63 supercell structure. Co on Cu (111) was modeled using a
3× 4 supercell of a Cu (111) surface with a thickness of 5 atomic layers and
a Co adatom on the surface, see Figure 5.6. All structures were relaxed until
the forces acting on each atom were below 0.02 eVA−1. For Co in bulk Co the
entire supercell and for Co on Cu (111) the adatom and the three topmost
Cu layers were relaxed. The PAW basis sets provide intrinsically projections
onto localized atomic orbitals, which we used to extract the hybridization
functions:
∆α1α2(ω) =
∑
k
V ∗kα1Vkα2
iω − ǫk. (5.1)
74
Transition metal impurities on metal surfaces
from our DFT calculations as outlined in Section 2.4.1.
The impurity model (2.9) can be solved without approximations using the
CTQMC technique. Both the weak-coupling [41] and hybridization expan-
sion [42, 43] algorithms can treat multi-orbital systems with general four-
fermion interaction terms. However, a strong-coupling approach is advan-
tageous in the case of a strongly interacting five orbital model, because the
expansion in the hybridization leads to much lower perturbation orders than
an expansion in the various interaction terms. For the interaction parame-
ters of the Co impurities used in this study the order in the hybridization
expansion was found to be a factor 20 lower compared to a weak coupling
expansion [81]. Furthermore, if the hybridization function is diagonal in the
orbital indices, as it is to a good approximation the case in the Co/Cu sys-
tems studied here, no sign problem appears, in contrast to the weak-coupling
approach, where correlated hopping terms were found to lead to severe sign
cancellations [81].
5.3.3 Results
The DFT calculations yield the orbital dependent hybridization functions
shown in Figure 5.7. In bulk Cu, the environment of the Co impurities is
cubically symmetric and the hybridization function decomposes into three-
fold degenerate t2g and twofold degenerate eg blocks. In the bulk symmetry
forbids off-diagonal elements in the hybridization function. On the surface,
the symmetry is reduced to C3v. For Co on Cu the hybridization function
decomposes into two twofold degenerate blocks transforming according to the
E-irreducible representation of C3v (E1 (dxz, dyz) and E2 (dx2−y2 , dxy)) and
the dz2-orbital transforming according to the A1 representation. For Co on
Cu the hybridization functions contain small off-diagonal matrix elements.
These off-diagonal elements, which are smaller than the diagonal ones, will
be neglected in our simulations. As a general trend, one can already see that
the hybridization of the Co d-electrons is about twice larger in the bulk than
on the surface.
DFT calculations are also used to calculate the occupancy of the Co 3d
impurity orbitals. To this end, we performed spin-polarized DFT calculations
using GGA as well as GGA+U of Co in and on Cu with the full interaction
vertex defined via the average screened Coulomb interaction U = 4 eV and
the exchange parameter J = 0.9 eV. We obtained the occupancies of the Co
75
5.3 Kondo physics in Co on Cu and Co in Cu
a) b)
−1.4
−1.2
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
−8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5
e d+
Re
∆ (e
V)
E (eV)
Co in Cu
egt2g
−1.4
−1.2
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
−8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5
e d+
Re
∆ (e
V)
E (eV)
Co on Cu
E2E1A1
−2
−1.8
−1.6
−1.4
−1.2
−1
−0.8
−0.6
−0.4
−0.2
0
−8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5
Im ∆
(eV
)
E (eV)
Co in Cu
egt2g
−1.4
−1.2
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
−8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5
Im ∆
(eV
)
E (eV)
Co on Cu
E2E1A1
Figure 5.7: Hybridization functions and crystal fields for Co in bulk Cu (a)
and as ad-atom on Cu (111) (b). In the upper panel the dynamical crystal field
ǫd + Re∆ is shown. Lower panel: Im∆.
3d orbitals derived from the PAW projectors n = n↑ + n↓ and the impurity
spin Sz =12(n↑ − n↓) and present them in Tab. 5.1.
In all cases the average Co 3d occupancy suggested by our DFT calculations
is between n = 7 and n = 8. For Co on Cu, the impurity spin is Sz ≈ 1
which is well in line with a d8 configuration of the Co. In the bulk, the Co
spin is Sz ≈ 1 in GGA+U and Sz ≈ 1/2 in GGA.
In the following we study Co in and on Cu in the five-orbital Anderson impu-
rity model formulation Eq. (2.9). In this framework, the chemical potential
has to be chosen to fix the occupancy of the Co d-orbitals. Due to the well
know double-counting problem in LDA+DMFT type approaches [82], the
precise chemical potential µ and the Co d-occupancy are not known. There-
76
Transition metal impurities on metal surfaces
GGA GGA+U
n Sz n Sz n Sz n SzCo on Cu 7.3 0.96 7.6 1.00 7.4 0.96 7.7 1.00
Co in Cu 7.3 0.51 7.6 0.53 7.3 0.90 7.5 0.93
Table 5.1: Occupancies n and impurity spins S as obtained from our GGA and
GGA+U calculations. Values obtained directly from the PAW projectors (n, S)
and normalized by the integrated total Co d-electron DOS, N =∫ν(E)dE, are
shown (n = n/N , Sz = Sz/N ).
fore, we computed results in a range of chemical potential values which yield
a total d occupancy consistent with the estimates of the DFT calculations.
For both systems the results of the DFT calculations predict a total den-
sity n . 8 and suggest a spin S ≈ 1 or slightly below in the case of Co in
Cu. For µ = 26, 27, 28 eV (Co in Cu) and µ = 27, 28, 29 eV (Co on Cu) we
obtain total densities and spins close to these DFT estimates. The values
of both observables for the lowest simulation temperature T = 0.025 eV are
presented in Table 5.2.
Table 5.2: Total density and spin.
System µ (eV) 〈n〉 〈S〉Co in Cu 26 7.51± 0.07 1.02± 0.02
Co in Cu 27 7.78± 0.05 0.92± 0.02
Co in Cu 28 8.06± 0.03 0.817± 0.007
Co on Cu 27 7.76± 0.05 1.07± 0.01
Co on Cu 28 7.93± 0.05 0.99± 0.01
Co on Cu 29 8.21± 0.03 0.860± 0.007
5.3.3.1 Quasiparticle spectra
We now analyze the excitation spectra of the Co impurities in order to under-
stand the dominant physics at different energy scales. For a first, qualitative
insight into the strength of many body renormalizations, we compare in Fig-
ure 5.8 the Co 3d-electron DOS obtained from our DFT calculations to the
Co 3d spectral functions obtained from analytical continuation of our QMC
results (Figure 5.8).
The non-spinpolarized GGA calculations used to determine the hybridization
functions yield — by definition — the LDOS corresponding to the Anderson
77
5.3 Kondo physics in Co on Cu and Co in Cu
0
0.2
0.4
0.6
0.8
1
1.2
−6 −5 −4 −3 −2 −1 0 1 2 3 4
DO
S
E (eV)
Co in Cu
GGAGGA+Uµ=26eVµ=27eV
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
−6 −5 −4 −3 −2 −1 0 1 2 3 4
DO
S
E (eV)
Co on CuGGA
GGA+Uµ=28eVµ=29eV
Figure 5.8: DOS of the Co impurities in bulk Cu (left) and on Cu (111) (right)
obtained from DFT (GGA and GGA+U) as well as QMC simulations at temper-
ature T = 0.025 eV. QMC results obtained at different chemical potentials µ are
shown.
model without two-particle interactions (U = J = 0 eV). For both, Co in and
on Cu, the GGA DOS exhibits a peak near the Fermi level (EF = 0). The
QMC DOS qualitatively reproduces the GGA DOS for the case of Co in Cu.
Here, the main difference between the two approaches is that QMC yields a
peak near the Fermi level which is approximately twice narrower and shifted
towards EF . GGA+U accounts for the local Coulomb interactions at the Co
atoms on a Hartree Fock level which leads to the destruction of the quasi-
particle peak near EF with all the spectral weight shifted to broad Hubbard
bands. The comparison to the QMC results shows that this destruction of
the quasi-particle peak is unphysical.
For Co on Cu the hybridization is weaker and the DOS from the QMC
simulations exhibits both quasi-particle peaks near EF as well as Hubbard
type bands at higher energies. The reduction of spectral weight of the quasi-
particle peak as compared to GGA is stronger here.
The orbitally resolved DOS of Co in and on Co is shown in Figure 5.9. For
Co in Cu the DOS of the eg and the t2g orbitals is very similar particularly
regarding the quasi-particle peak — despite the (energy dependent) crystal
field splitting on the order of some 0.1 eV.
The DOS of Co on Cu exhibits stronger differences between the E1, E2, and
A1 orbitals. The E2 orbitals, which spread out perpendicular to the z-axis,
show the weakest hybridization effects, but even here, a quasi-particle peak
78
Transition metal impurities on metal surfaces
0
0.2
0.4
0.6
0.8
1
1.2
−6 −5 −4 −3 −2 −1 0 1 2 3 4
DO
S
E (eV)
Co in Cu
egt2g
0
0.2
0.4
0.6
−6 −5 −4 −3 −2 −1 0 1 2 3 4
DO
S
E (eV)
Co on Cu
E2E1A1
Figure 5.9: Orbitally resolved DOS of the Co impurities in bulk Cu (left) and on
Co (111) (right) obtained from our QMC simulations at temperature T = 0.025 eV
and chemical potential µ = 27 eV and µ = 28 eV, respectively.
appears in all orbitals. The appearance of low energy quasi-particle peaks in
all orbitals is different from the behavior expected for a spin-1 two-channel
Kondo model, where a low energy quasi-particle resonance would be observed
in two orbitals (four spin-orbitals) only.
The DOS as obtained from our QMC calculations suggests a low-temperature
Fermi liquid state involving all orbitals for both, Co in and on Cu. We
investigate the nature of this state in the following sections by analyzing the
self-energies obtained from QMC and the statistics of relevant atomic states.
5.3.3.2 Low energy Fermi liquid
If a Fermi liquid develops, the self-energy takes the form
Σ(T, ω) = Σ(T, 0) + Σ′(T, 0)ω +O(ω2) (5.2)
with Σ(T, 0) and the first energy derivative Σ′(T, 0) being real for T → 0. In
this regime, the spectral weight Z associated with the quasi-particle peak is
determined by
Z = (1− ReΣ′(0))−1. (5.3)
Our QMC calculations yield the self-energy on the Matsubara axis. Analytic
continuation ω → iωn shows that Fermi liquid behavior manifests itself on
the Matsubara axis by
ImΣ(T, iωn) ≈ ImΣ(T, 0)− ImΣ′(T, 0)ωn (5.4)
79
5.3 Kondo physics in Co on Cu and Co in Cu
at low frequencies with ImΣ(T, 0) ∼ T 2.
We now compare these relations to the frequency and the temperature depen-
dence of ImΣ(T, iωn) obtained from our QMC calculations. At the lowest ac-
cessible temperature, T = 0.025 eV, we obtained the Matsubara self-energies
depicted for Co in and on Cu in Figure 5.10. In both systems |ImΣ| clearlydecreases as ωn → 0, for all orbitals except for the E2 orbitals of Co on Cu at
µ = 27 eV. This is clearly different from the diverging Σ(iωn) ∼ 1iωn
, expected
for the localized moment of an isolated atom. For Co in bulk Cu, the eg and
t2g orbitals exhibit very similar self-energies, whose low-energy behavior is
consistent with the form expected for a Fermi-liquid (Eq. 5.4). For Co on Cu,
the self-energies differ considerably between the different orbitals with the E1
orbitals being least correlated and the E2 orbitals exhibiting the largest self-
energies at low frequencies. Our results indicate that a Fermi liquid develops
in all Co orbitals, also here, although the Kondo temperature appears to be
orbital dependent.
5.3.3.3 Estimation of TK from QMC
To define a Kondo temperature scale even in cases without well defined local
moment at intermediate temperatures, we define TK through the width of
the quasi-particle resonance in the single particle spectral function near EF ,
which is measured in STM experiments.
In our QMC simulations we determine TK from the quasi-particle weight
Z. The simulations yield the self-energy at the Matsubara frequencies ωn =(2n+1)π
β. Analytical continuation of Eq. (5.3) yields
Z ≈(
1− ∂ImΣ(iωn)
∂iωn
∣∣∣∣ωn=0
− Re∆′(0)
)−1
(5.5)
and we use Eq. (5.4) to evaluate the derivative. In Figure 5.11 we show the
quasi-particle weight of Co in Cu and Co on Cu for the different types of
orbitals as a function of the chemical potential µ. The values of degenerate
orbitals agree within an accuracy of 10−2 to which precision we are listing
them in Tab. 5.3 for T = 0.025 eV. The systems whose spin is closest to
S = 1 are found to have the lowest values of Z. Co in Cu clearly has higher
quasi-particle weights compared to Co on Cu. As we will see, this results
in a higher Kondo temperature TK , which is also confirmed experimentally.
In experiments using STM measurements the Kondo temperature has been
80
Transition metal impurities on metal surfaces
−1
−0.8
−0.6
−0.4
−0.2
0
0 1 2 3 4 5 6
Im Σ
(eV
)
ωn (eV)
Co in Cueg
t2g
−1.8
−1.6
−1.4
−1.2
−1
−0.8
−0.6
−0.4
−0.2
0
0 1 2 3 4 5 6Im
Σ (
eV)
ωn (eV)
Co on CuE2E1A1
−1
−0.8
−0.6
−0.4
−0.2
0
0 1 2 3 4 5 6
Im Σ
(eV
)
ωn (eV)
Co in Cueg
t2g
−1.8
−1.6
−1.4
−1.2
−1
−0.8
−0.6
−0.4
−0.2
0
0 1 2 3 4 5 6
Im Σ
(eV
)
ωn (eV)
Co on CuE2E1A1
−1
−0.8
−0.6
−0.4
−0.2
0
0 1 2 3 4 5 6
Im Σ
(eV
)
ωn (eV)
Co in Cueg
t2g
−1.8
−1.6
−1.4
−1.2
−1
−0.8
−0.6
−0.4
−0.2
0
0 1 2 3 4 5 6
Im Σ
(eV
)
ωn (eV)
Co on CuE2E1A1
Figure 5.10: Orbitally resolved self-energies for Co in Cu (left panel) and Co
on Cu (right panel). From top to bottom , ImΣ(iω) is shown for the chemical
potentials µ = 26, 27, 28 eV (Co in Cu) and µ = 27, 28, 29 eV (Co on Cu) .
81
5.3 Kondo physics in Co on Cu and Co in Cu
0
0.2
0.4
0.6
0.8
1
25 26 27 28 29 30
Z
µ
egt2gE2E1A1
Figure 5.11: Orbitally resolved quasi-particle weight of Co in Cu and Co on Cu
at temperature T = 0.025 eV.
found to be TK = 655K ± 155K = 0.056 ± 0.013 eV for Co in Cu [74] and
TK ≈ 54± 5K = 0.0046± 0.0005 eV for Co on Cu [68, 71, 74, 80].
Following Hewson’s derivation of a renormalized perturbation theory of the
Anderson model [83] we use as a definition for the Kondo temperature:
TK = −π4ZIm∆(0). (5.6)
The values computed according to Eq. (5.6) are listed in Tab. 5.3 for all
impurity energy levels Eα+µ at temperature T = 0.025 eV. As for the quasi-
particle weights we find the lowest values of TK for µ = 26 eV (Co in Cu) and
µ = 28 eV (Co on Cu). Averaging over orbitals we obtain TK = 0.118 eV (Co
in Cu, µ = 26 eV, T = 0.025 eV) and TK = 0.016 eV (Co on Cu, µ = 28 eV,
T = 0.025 eV) with a ratio of T INK /TON
K = 7.4. This large difference between
the two Kondo temperatures is in fair agreement with experiments, where a
ratio of T INK /TON
K = 12.1 has been found [74]. As discussed in Section 5.3.4 the
physical quantities which determine the Kondo temperature scale enter in the
argument of an exponential function, suggesting that a comparison of the log-
arithms log(Texp)/ log(TK) is more appropriate when judging the predictive
82
Transition metal impurities on metal surfaces
power of our first-principles simulations. We find log(Texp)/ log(TK) = 1.4
(Co in Cu, µ = 26 eV, T = 0.025 eV) and log(Texp)/ log(TK) = 1.3 (Co on
Cu, µ = 28 eV, T = 0.025 eV).
Table 5.3: Kondo temperatures TK computed from the quasi-particle weight Z
and the hybridization function ∆(ω) according to Eq. (5.6) at the lowest sim-
ulation temperature T = 0.025 eV. The experimental Kondo temperatures are
TK = 0.056±0.013 eV (Co in Cu) and TK = 0.0046±0.0002 eV (Co on Cu) [68,74]
.
System µ (eV) Ei -Im(∆(0)) (eV) Z TK (eV)
Co in Cu 26 -0.288 0.43± 0.01 0.38 0.13
Co in Cu 26 -0.44 0.340± 0.009 0.39 0.10
Co in Cu 27 -0.288 0.43± 0.01 0.42 0.14
Co in Cu 27 -0.44 0.340± 0.009 0.47 0.12
Co in Cu 28 -0.288 0.43± 0.01 0.48 0.16
Co in Cu 28 -0.44 0.340± 0.009 0.56 0.15
Co on Cu 27 -0.12 0.124± 0.002 0.06 0.006
Co on Cu 27 -0.39 0.226± 0.002 0.19 0.03
Co on Cu 27 -0.34 0.197± 0.001 0.08 0.01
Co on Cu 28 -0.12 0.124± 0.002 0.05 0.005
Co on Cu 28 -0.39 0.226± 0.002 0.15 0.03
Co on Cu 28 -0.34 0.197± 0.001 0.10 0.01
Co on Cu 29 -0.12 0.124± 0.002 0.14 0.01
Co on Cu 29 -0.39 0.226± 0.002 0.26 0.05
Co on Cu 29 -0.34 0.197± 0.001 0.27 0.04
For both systems our computed Kondo temperatures are higher than the
experimentally determined values. This can be either due to the neglect
of spin-orbit coupling effects which lifts degeneracies and narrows the low
energy resonances or due to the Coulomb interactions being larger than the
U = 4 eV assumed here.
The temperature dependence of the self-energy allows for an alternative test
whether and at which energy scale a Fermi liquid emerges. This is illustrated
for Co in and on Cu in Figure 5.12. We find an almost temperature indepen-
dent behavior of ImΣ(T, iωn) for Co in Cu if T < 0.05 eV, which provides an
estimate of TK ≈ 0.05 eV. In the case of Co on Cu ImΣ(T, iωn) still evolves
as one lowers the temperature from T = 0.05 eV to T = 0.025 eV, which
suggests a lower Kondo temperature.
According to Eq. (5.4), the linear extrapolation of ImΣ(T, iω) to iωn → 0
83
5.3 Kondo physics in Co on Cu and Co in Cu
−1.4
−1.2
−1
−0.8
−0.6
−0.4
−0.2
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Im Σ
(eV
)
ωn (eV)
Co on Cu
T=0.05eVT=0.025eV
−1
−0.8
−0.6
−0.4
−0.2
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Im Σ
(eV
)ωn (eV)
Co in Cu
T=0.2eVT=0.1eV
T=0.05eVT=0.025eV
−1.4
−1.2
−1
−0.8
−0.6
−0.4
−0.2
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Im Σ
(eV
)
ωn (eV)
Co on Cu
T=0.05eVT=0.025eV
−1
−0.8
−0.6
−0.4
−0.2
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Im Σ
(eV
)
ωn (eV)
Co in Cu
T=0.2eVT=0.1eV
T=0.05eVT=0.025eV
−1.4
−1.2
−1
−0.8
−0.6
−0.4
−0.2
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Im Σ
(eV
)
ωn (eV)
Co on Cu
T=0.05eVT=0.025eV
Figure 5.12: Left panel: Im(Σ) of Co on Cu for µ = 29 eV at orbital energies
Ei = −0.12 eV (dxy, dx2−y2), Ei = −0.39 eV (dyz , dxz) and Ei = −0.34 eV (dz2).
Right panel:Im(Σ) of Co in Cu for µ = 28 eV at orbital energies Ei = −0.288 eV(dxy, dyz , dxz) and Ei = −0.44 (dz2 , dx2−y2).
84
Transition metal impurities on metal surfaces
−1
−0.8
−0.6
−0.4
−0.2
0
0 0.05 0.1 0.15 0.2
Inte
rcep
t
T (eV)
ImΣ → aωn+bfit of Im Σ(0,T), t2g
−1
−0.8
−0.6
−0.4
−0.2
0
0 0.05 0.1 0.15 0.2
Inte
rcep
t
T (eV)
ImΣ → aωn+bfit of Im Σ(0,T), eg
Figure 5.13: Temperature dependent quasi-particle lifetimes for Co in Cu at
µ = 26 eV for the t2g (left) and eg orbitals (right).
yields the inverse lifetime ~τ−1 = ImΣ(T, 0) of quasi-particles at the Fermi
level if a Fermi liquid is formed. In Figure 5.13 we plot this extrapolation
for Co in Cu, µ = 26 eV and show ImΣ(T, 0) as a function of temperature.
We find a ImΣ(T, 0) ∼ T 2-behavior for both sets of orbitals, which corrobo-
rates the formation of a Fermi-liquid in the eg and t2g orbitals at the lowest
accessible temperatures.
5.3.4 Discussion
Our QMC calculations showed that, for both Co in and on Cu Fermi liquids
involving all orbitals of the Co impurity form at low temperatures. We now
want to understand these results on the basis of scaling arguments starting
with (higher energy) charge fluctuations going to (lower energy) spin and
orbital fluctuations.
5.3.4.1 Charge fluctuations
With our values of the Coulomb interaction strength in the local Hamiltonian
(U = 4 eV; J = 0.9 eV) the energies of removing (E−) or adding (E+) an elec-
tron to the impurity are E− ≈ E+ ≈ 2 eV. A first order expansion in the hy-
bridization gives a qualitative estimate of the role of charge fluctuations [75]:
The norm of the admixtures of d7 and d9 configurations to a predominantly
d8 ground state of the impurity is approximately Nn 6=8 = − 1πIm∆(0) 10
U/2. For
Co in Cu, −Im∆(0) ≈ 0.4 eV leads to Nn 6=8 ≈ 0.6. The hybridization of Co
85
5.3 Kondo physics in Co on Cu and Co in Cu
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 1 2 3 4 5 6 7 8 9 10
Pro
bab
ilit
y
Occupation
Co in Cu
µ=26eVµ=27eVµ=28eV
GGA
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 1 2 3 4 5 6 7 8 9 10
Pro
bab
ilit
y
Occupation
Co on Cu
µ=27eVµ=28eVµ=29eV
GGA
Figure 5.14: Left: Occupation statistics of Co in Cu for T = 0.025 eV . Right:
Occupation statistics of Co on Cu for T = 0.025 eV.
on Cu is about twice smaller, −Im∆(0) ≈ 0.2 eV, yielding a correspondingly
smaller weight of non-d8 configurations Nn 6=8 ≈ 0.3.
Our QMC calculations allow to quantitatively measure the charge fluctua-
tions. In Figure 5.14 we plot the “valence histogram” [84] for Co in/on Cu
for different choices of the chemical potential and compare it to the “valence
histogram” of the Slater determinant built from the lowest GGA eigenstates.
The histogram shows the weights which the eigenstates in the different charge
sectors n = 0, 1, . . . 10 contribute to the partition function. In all cases the
local Coulomb interaction in the QMC simulation leads to a narrower dis-
tribution of the occupancies as compared to the GGA valence histograms.
This effect is most pronounced in the case of Co on Cu (111), where the d8
configuration clearly dominates over the d7 and d9 configurations. For Co in
Cu, there are still noticeable correlations and the narrowing of the valance
histogram as compared to the GGA case but the d8 configuration contributes
only about 50% in the QMC simulations, with significant weight coming from
the d7, d9 and even the d6 and d10 configurations. The measured values for
Nn 6=8 (≈ 0.5 for Co in Cu and ≈ 0.3 for Co on Cu) agree surprisingly well
with the simple perturbative estimate.
A Schrieffer-Wolff decoupling of ionized impurity states to discuss the low
energy physics can be justified if the weight of non-d8 configurations is
Nn 6=8 ≪ 1. Hence, an intermediate energy scale of well formed (unscreened)
fluctuating spin or orbital moments at frozen impurity valency might be de-
fined in the case of Co on Cu (111) but clearly not in the case of Co in
86
Transition metal impurities on metal surfaces
Cu.
5.3.4.2 Spin and orbital fluctuations
While for Co in Cu charge, spin and orbital fluctuations will be present
down to lowest energies, for Co on Cu only fluctuations of the orbital and
the spin degree of freedom are expected to dominate the low energy physics.
To further investigate to which extent orbital and spin fluctuations might
determine the low energy behavior of the impurity we estimate Kondo tem-
peratures within simplified models and compare these estimates to our QMC
calculations as well as experiments.
Assuming well-defined magnetic moments (Nn 6=8 ≪ 1) a scaling analysis
(cf. Ref. [75]) allows to estimate the Kondo scale analytically in simplified
situations: If neither Hund’s rule coupling nor crystal field splitting or any
other symmetry breaking terms were present the spin and the orbital degrees
of freedom of the Co impurities could fluctuate freely and independently.
This would lead to a Kondo temperature [75] TK ∼ D0 exp [−1/2Nn 6=8],
where D0 ∼ min(E±,Λ) is related to the impurity charging energies (E±)
and the electronic bandwidth Λ. D0 is on the order of several eV. With
D0 = U/2 = 2 eV we estimate TK ≈ 0.4D0 = 0.9 eV in the case of Co in Cu
and TK ≈ 0.2D0 = 0.4 eV in the case of Co on Cu. This is in both cases (at
least) an order of magnitude larger than the Kondo temperatures obtained
from QMC and the experimentally measured Kondo temperatures.
The opposite limit is given by the case with strong Hund’s rule coupling
and supressed orbital fluctuations. Without orbital fluctuations, but still
disregarding the Hund’s coupling J , the Kondo temperature reads [66, 75]
TK ∼ D0 exp[
πU
8Im∆(0)
]
= D0 exp[
− 52Nn6=8
]
, which would lead to TK =
0.02D0 ≈ 0.04 eV for Co in Cu. The Hund’s rule coupling reduces the
Kondo temperature [77] further to T ∗K = TK(TK/JHS)
2S−1. With S = 1
and JH = 0.9 eV this would lead to T ∗K ≈ 0.002 eV. For Co on Cu, the as-
sumption of an orbital singlet yields TK = 0.0004D0 ≈ 0.001 eV and Hund’s
rule coupling further reduces the Kondo temperature to T ∗K ≈ 1µ eV. This
limit thus yields Kondo temperatures which are orders of magnitude smaller
than those obtained in our QMC calculations as well as the Kondo temper-
atures measured experimentally for Co in and on Cu.
It is thus the successive locking of the impurity electrons to a large spin by the
Hund’s rule coupling and the partial freezing out of orbital fluctuations that
87
5.3 Kondo physics in Co on Cu and Co in Cu
0
0.2
0.4
0.6
0.8
1
−1 −0.5 0 0.5 1
DO
S
E (eV)
J=0.9eVJ=0.0eV
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
28.4 28.6 28.8 29
Z
µ+Eα (eV)
E2
A1E1
J=0.9eVJ=0.0eV
Figure 5.15: Comparison of the DOS (left) and quasi-particle weight Z (right)
of Co on Cu at U = 4 eV for the two values J = 0.0 eV and J = 0.9 eV.
determines the on-set of Fermi liquid behavior and the Kondo temperature
in realistic systems like Co in or on Cu.
With this in mind it is instructive to analyze the influence of a static crystal
field on the energy spectrum of otherwise isolated Co atoms. Without crystal
fields, in a d8 configuration our local Coulomb interaction (U = 4 eV; J =
0.9 eV) yields a 21 fold degenerate L = 3, S = 1 ground state which is
separated from the L = 2, S = 0 multiplet by an energy of EL=2,S=0 = 1.3 eV.
This is clearly larger than the crystal field acting on the Co impurities (Figure
5.7): The cubic crystal field (evaluated at the Fermi level) of Co in Cu leads to
the eg states being 0.18 eV higher in energy than the t2g states. In this crystal
field, the resulting d8 ground state is an orbital singlet. Excitations to higher
crystal field split states require energies on the order of 0.2 eV. This is larger
but comparable in order of magnitude to the Kondo temperatures obtained
in experiments and simulation. However, fluctuations to these higher crystal
field split states must be taken into account to explain the characteristic
temperature of the low energy Fermi liquid formed at Co impurities in Cu.
In our model of Co on Cu, the static crystal fields also lift the degeneracy
of the ground state multiplet but a double degeneracy in the orbital space
remains. Excitations to higher crystal field split states require 0.03−0.08 eV.In this model, even the ground state multiplet allows for fluctuations of the
orbital degree of freedom.
In order to examine the effect of constraining orbital fluctuations we consider
the case of Co on Cu which exhibits a strong reduction of the quasi-particle
88
Transition metal impurities on metal surfaces
peak compared to the GGA spectral function representing the U = 0, J = 0
case. Turning off the Hund’s coupling J allows the orbital and spin degrees of
freedom to fluctuate more freely and given the scaling considerations should
result in a higher Kondo temperature as well as a broader quasi-particle peak.
We study the effect of J = 0 for Co on Cu, T = 0.025 eV, µ = 29 eV and
present the comparison of the quasi-particle weight and peak in Figure 5.15.
In line with our statement that the Kondo temperature is determined by the
locking of the impurity electrons to a larger spin and possible restrictions of
the orbital fluctuations, we find a broadening of the quasi-particle peak and
increase of the quasi-particle weight Z as J → 0.
5.3.5 Conclusions
For Co in and on Cu we found that a Fermi liquid is formed at low T in-
volving all impurity d-orbitals. The example of Co on Cu shows that the
characteristic temperature, TK , associated with the onset of Fermi liquid be-
havior can differ between the impurity orbitals. The comparison of our QMC
calculations and scaling arguments further demonstrates that fluctuations in
the orbital degree of freedom and Hund’s rule coupling are crucial in deter-
mining TK in realistic systems. This is well beyond the physics of simple
“spin-only” models. The understanding of magnetic nanostructures based
on 3d adatoms on surfaces as well as magnetic impurities in bulk metals re-
quire us to account for the orbital degrees of freedom. DMFT provides a link
between quantum impurity problems and extended lattices of atoms which
are subject to strong electron correlations.
It remains thus a future challenge to understand how the orbital degree
of freedom controls the quenching of magnetic moments and eventually the
formation of low energy Fermi liquids in realistic extended correlated electron
systems as well as magnetic nanostructures.
89
5.3 Kondo physics in Co on Cu and Co in Cu
90
Chapter 6
Dynamical Mean Field Theory
U
↑↓
t
Time
Electron Reservoir
Vpe−
V ∗p
e−
Figure 6.1: Sketch of the DMFT-mapping: A site within the lattice is singled
out and described as a quantum impurity system. The dynamics on the impurity
is adjusted to mimic the lattice system.
6.1 Introduction
The Hubbard model has been proposed as a relatively simple model to de-
scribe strongly correlated systems. Despite its simplicity the Hubbard model
is very difficult to solve analytically as well numerically in two and three
dimensions and requires approximate methods. It has been first realized by
Muller-Hartmann [85] and Metzner and Vollhardt [86] that its diagrammat-
ics simplifies in the limit of infinite dimension. Later, Georges, Kotliar and
coworkers showed that the solution of this quantum many body problem
can be formulated as the solution of a quantum impurity system subject to a
91
6.1 Introduction
self-consistency condition which connects the impurity problem to the lattice
problem [87–89].
The key concept of the DMFA is sketched in Figure 6.1. In DMFA the lattice
model is mapped onto a single-site problem and a self-consistency condition.
The idea can be well explained by considering the classical analogy of the
Ising model. In the mean-field theory of the Ising model we map the lattice
model:
H = −J∑
〈ij〉
SiSj (6.1)
with coordination number z onto a single-site effective model describing a
single spin S0 in an effective field heff = Jzm
Heff = −heffS0. (6.2)
The spin degrees of freedom of the neighboring lattice sites have therefore
been replaced by an effective field heff. Within this model we can easily
evaluate the effective magnetization meff:
meff = tanh(βheff) (6.3)
and we connect back to the lattice model by identifying the magnetization
m per site with meff:
m!= meff = tanh(βJzm) (6.4)
Above identification imposes a condition for the magnetization with the aim
to obtain a magnetization which is consistent with the lattice model.
Analogously, we single out a lattice site in the DMFA and couple it to a bath
of non-interacting fermions. The Hubbard model
H =∑
ijαβ
tαβij d†iασdjβσ +
∑
iαβγδσσ′
Uσσ′
αβγδd†iασd
†iβσ′diδσ′diγσ −
∑
iασ
µd†iασdiασ (6.5)
is thereby mapped onto a quantum impurity model:
H =∑
p
ǫpa†pap +
∑
pασ
V αp d
†ασap + V α∗
p a†pdασ
+∑
ασ
ǫαd†ασdασ +
∑
iαβγδσσ′
Uσσ′
αβγδd†iασd
†iβσ′diδσ′diγσ
(6.6)
and we enforce the electron transitions from and to the impurity to behave as
the electron motion on the lattice by identifying the lattice Green’s function
92
Dynamical Mean Field Theory
with the impurity Green’s function:∑
k
G(k, ω)!= Gimp(ω) (6.7)
The bath will thus mimic the electron motion on the lattice.
Also within DFT we encountered a similar mapping strategy. In DFT we map
a many-body problem onto a single-effective particle problem and adjust the
Hartree and exchange-correlation potential consistently with the density. It
is therefore not astonishing that we can phrase the self-consistency conditions
to solve the Hubbard model within a Functional Formalism as well.
In analogy to the Hohenberg-Kohn functional the Luttinger Ward-Functional
[88]
Φ [G] = Φuniv [G] + Trln [G]− Tr[G−1
0 G]
(6.8)
is a functional of the Green’s function, where G0 is the Green’s function of
the non-interacting reference system. In this non-interacting reference system
(U = 0) the Greens function is given by
G0(k, ω) =1
ω + µ− ǫk. (6.9)
The effects of the interaction are encoded in the self-energy Σ(k, ω) and the
Greens function of the interacting model is given by
G(k, ω) =1
ω + µ− ǫk − Σ(k, ω)(6.10)
The Dyson equation relates G and G0 to the self-energy:
Σ(k, ω) = G0(k, ω)−1 −G(k, ω)−1. (6.11)
Similar to the Hohenberg Kohn functional the Luttinger Ward functional
Φ [G] is extremal at the physical correct Green’s function. As in DFT the
functional has a universal part Φuniv and a material specific part. Since Φuniv
is the sum of all vacuum to vacuum diagrams the functional derivative:
δΦuniv [G]δG
= Σ (6.12)
delivers the self-energy, which is the conjugate observable of the Green’s
function. Using a Legendre transform the functional can also be expressed
as a functional of the self-energy:
Φ [Σ] = Φuniv [Σ]− Trln[G−1
0 − Σ]
(6.13)
93
6.2 Self-consistency equations
Motivated by the observation that the self-energy becomes momentum-independent
in the limit of infinite coordination number we replace the self-energy Σ(k, ω)
by a momentum independent self-energy Σ(ω) and arrive at an approximate
Luttinger-Ward functional:
Φapprox [Σapprox] = Φapproxuniv [Σapprox]− Trln
[G−1
0 − Σapprox]. (6.14)
Even though this approximation is a priori uncontrolled, it is possible to
control this approximation by reintroducing the momentum dependence with
a small parameter [90, 91].
Since there is no spatial information encoded in Σ(ω) the functional deriva-
tive:δΦuniv [Σapprox]
δΣapprox= Gimp (6.15)
yields a Green’s function describing a quantum impurity.
The functional is stationary at the Dyson equation:
Σ−G−10 + G−1 = 0 (6.16)
where in DMFA we now provide the Green’s function by solving a quantum
impurity model.
6.2 Self-consistency equations
For the first iteration step we provide an initial guess for the hybridization
function ∆
∆(iωn) =∑
p
V 2p
iωp − ǫp(6.17)
or bath Green’s function G0
G0 = (iωn + µ−∆(iωn))−1 (6.18)
and simulate the Quantum impurity models defined in Section 2.2 using one
of the Impurity solvers presented in Chapter 3.
In the case of the CTQMC impurity solver we will compute the impurity
Green’s function in imaginary time:
Gimp(τ) = −〈Tτd(τ)d†(0)〉 (6.19)
94
Dynamical Mean Field Theory
Σ(iωn) = iωn + µ−∆(iωn)−G−1(iωn)
G(iωn) =∑
k
(iωn + µ− ǫ(k)− Σ(iωn)
)−1
∆(iωn) = iωn + µ− Σ(iωn)− G−1(iωn)
∆(iωn)
∆(τ)
G(τ)
G(iωn)
momentum-average
Dyson
Fourier transform
Impurity Solver
Fourier transform
Dyson
momentum-average
DMFT
Self-Consistency Loop
Figure 6.2: Iterative procedure of a DMFT-simulation. The impurity solver on
the left hand side provides the impurity Green’s function G(τ). Applying the self-
consistency equations on the right hand side yields a new hybridization function
∆(τ). The cycle is repeated until convergence is reached.
By Fourier transforming Gimp to the Matsubara frequency domain we can
extract the self-energy Σimp using the Dyson equation:
Σimp(iωn) = G−10 (iωn)−G−1
imp(iωn) (6.20)
In the DMFA we approximate the lattice self-energy Σ(k, iωn) by the mo-
mentum independent self-energy Σ(iωn) using Eq. (6.16). With Σ(k, iωn)←Σ(iωn) the momentum averaged Green’s function of Eq. (6.7) reads:
G(iωn) =∑
k
1
iωn + µ− ǫ(k)− Σ(iωn)(6.21)
or written as an integral over the lattice DOS D(ǫ):
G(iωn) =
∫ ∞
−∞
D(ǫ)
iωn + µ− ǫ− Σ(iωn)(6.22)
Applying again the Dyson equation with G = G
G−10 (iωn) = Σimp(iωn) + G−1 (6.23)
95
6.2 Self-consistency equations
we obtain a new bath Green’s function or using Eq. (6.18) a new hybridiza-
tion function which enters again the impurity solver.
Figure 6.3: Sketch of the Bethe lattice.
For a Bethe lattice [92] shown in Figure 6.3 with lattice DOS
D(ǫ) =
√
4− (ǫ/t)2
2πt(6.24)
Eq. (6.22) simplifies to
G0(iωn) = iωn + µ− t2G(iωn). (6.25)
The iterative procedure of a DMFT self-consistency routine is summarized in
the diagram in Figure 6.2. Computationally most expensive is the solution
of the impurity problem.
The converged solution contains the information about the lattice in the sense
that the exchange of electrons between the impurity and the bath will emulate
the topology of the lattice. Within the DMFA the lattice behavior is however
only captured to some extent, e.g. spatial fluctuations which are known to
play an important role in low-dimensional systems are by construction - the
quantum impurity model we are dealing with is a zero-dimensional system -
not incorporated [90].
A converged solution is usually achieved within 10-20 iterations depending
on whether the system is close to a phase transition or deep within a phase.
The converged solution can be processed further. From the Green’s function
G(τ) the spectral function can be computed via the analytic continuation
methods presented in Chapter 4 and compared to experimental data such
as (I)PES spectra or the differential conductance of STM measurements as
explained in Section 2.5.
Application wise we will first present a DMFT study of a model-Hamiltonian
system, namely the 5-band Hubbard model on the Bethe lattice which is
96
Dynamical Mean Field Theory
presented in the next Section. Second, we discuss the extension to describe
realistic materials and present a study of the two transition metal monoxides
NiO and CoO.
97
6.3 5-band Hubbard model on the Bethe lattice
6.3 5-band Hubbard model on the Bethe lat-
tice
6.3.1 Introduction
The 5-band Hubbard model on the Bethe lattice provides a non material-
specific model and gives insight into generic features of strongly correlated
multi-orbital systems. In particular, it serves as a generic model of transition
metal compounds whose electronic behavior is determined by the d-orbital
states close to the Fermi level.
We investigate the effect of the Hund’s coupling J on the magnetic and
metal-insulator behavior of the 5-band Hubbard model on the Bethe lattice.
Using single-site DMFT we explore the occurring phases as a function of the
chemical potential for two cases of the Hund’s coupling.
We will not only describe the variety of phases we find but also show how
starting from the valence statistics we can identify the exchange mechanisms
underlying the magnetic phases.
6.3.2 Model
We study the 5-orbital Hubbard model Eq. (2.8) in the Slater-Kanamori
formulation Eq. (2.17) within the single-site DMFA and use the Krylov
implementation of the hybridization algorithm to solve the impurity problem.
For a given U and J the 5-orbital system is at half-filling for the chemical
potential µhalf = 4.5U − 10J [47]. For the scan at different µ we restrict
ourselves to the range of densities n ∈ [0, 5] since the model is particle-hole
symmetric. We explore the phase diagram for two Hund’s coupling values
J = U/4 and J = U/6 at a Coulomb interaction U = 12t and temperature
T = 0.02t.
Even though the study of magnetic ordering induced by spontaneous sym-
metry breaking in principle requires to simulate a quantum impurity model
for each sublattice, Chan et al. [93] showed that it is possible to study the
magnetic and orbital ordering by analyzing the iterative behavior of the
DMFT-cycles. In this study we applied this inexpensive method to estimate
two-sublattice order.
We use the self-consistency equation Eq. (6.25) to impose the self-consistency
condition on the Bethe lattice. It is often customary to enforce symmetries
98
Dynamical Mean Field Theory
within a DMFT self-consistency step and to symmetrize the Green’s and
hybridization function over degenerate spin-orbitals. Here, however, we do
not perform such a symmetrization and as a result encounter oscillations in
the DMFT solution as a function of iterations. The crucial point here is that
these oscillations indicate the presence of two-sublattice order which we use
to determine the magnetic phases.
6.3.3 Results
6.3.3.1 Determination of the phases
0
0.02
0.04
0.06
0.08
0.1
0 2 4 6 8 10 12 14 16 18
Orb
ital
Occ
upat
ion
Iteration
(a)
0
0.1
0.2
0.3
0.4
0.5
0 2 4 6 8 10 12 14 16 18
Orb
ital
Occ
upat
ion
Iteration
(b)
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10 12 14 16 18
Orb
ital
Occ
upati
on
Iteration
(c)
Figure 6.4: Orbital occupations as a function of DMFT-iterations. A paramag-
netic (a), ferromagnetic (b) and antiferromagnetic phase (c) can be distinguished.
99
6.3 5-band Hubbard model on the Bethe lattice
In Figure 6.4 we show the orbital occupation of the 10 spin-orbitals as a
function of DMFT iterations of three cases representing different magnetic
ordering. At the very low density n = 0.81 corresponding to a chemical
potential µ = 1t all the orbital occupancies fluctuate around ni ≈ 0.08
(Figure 6.4(a)). No specific pattern can be observed from which we infer
that the system is paramagnetic. In Figure 6.4(b) we show the example of
ferromagnetic ordering. It can be clearly seen how the orbital occupancies
for spin-down (blue) and spin-up (orange) start to differ in the first iteration
steps and stabilize at the two values ni ≈ 0.42 and ni ≈ 0.01. Since orbitals of
only one spin-species are dominantly occupied the system is magnetic. Also
the system presented in Figure 6.4(c) exhibits magnetic ordering. Here we see
how the orbital occupancies of different spin species oscillate between n = 1
and n = 0. As shown by Chan et al. [93] oscillations signal a two-sublattice
ordering and we conclude that the system is antiferromagnetic.
As can be seen from these three representatives the paramagnetic (PM),
ferromagnetic (FM) and antiferromagnetic (AFM) phases can be clearly dis-
tinguished by this analysis. In this way we inspected the DMFT-iterations
for all J and µ studied and present the densities associated with these phases
in Figure 6.5.
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
n(µ
)
µ
PMFM
AFM
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28
n(µ
)
µ
PMFM
AFM
Figure 6.5: Density as a function of chemical potential for J = U/4 (left) and
J = U/6 (right). The corresponding magnetic phase is indicated.
For both J = U/6 and J = U/4 we find paramagnetism for total densities
n < 1, which is in agreement with the three-orbital study by Chan et al. [93].
At half-filling the system is antiferromagnetic. According to Hund’s rule
the spins at a single site will arrange such that the total spin is maximized.
100
Dynamical Mean Field Theory
At half filling this implies that each orbital is occupied by one specific spin
orientation and due to the Pauli blocking electron exchange with the bath is
only possible for electrons with opposite spins. It is therefore favorable that
neighboring sites have opposite spin orientation and hence antiferromagnetic
order is favored. For J = U/6 we find antiferromagnetism for n = 4 also. At
other intermediate densities n ∈ (1, 5) we find the system to be ferromagnetic.
Compared to the three-orbital case ferromagnetism persists over a wider
range.
For the ferromagnetic order we additionally distinguish whether the system
is fully-polarized or not.
Close to the integer densities n=3,4 the ferromagnetic systems are fully-
polarized resulting in insulating orbitals for one of the spin-species, whereas
at lower densities in the ferromagnetic regime the system is not fully-polarized.
We find the fully-polarized regime to be sensitive to the data accuracy. In
a first attempt to explore the phase diagram with short simulation runs
we found the oscillatory behavior shown in the right panel of Figure 6.6
suggesting according to Ref. [93] a two-sublattice orbital ordering pattern.
However, this turned out to be an artifact, since the oscillatory behavior
is not reproduced when increasing the simulation time as shown in the left
panel of Figure 6.6.
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10 12 14 16 18
Orb
ital
Occ
upati
on
Iteration
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10 12 14 16 18
Orb
ital
Occ
upati
on
Iteration
Figure 6.6: Orbital occupations as a function of DMFT iterations in the fully
polarized regime. Long simulation runs (left) yield a ferromagnetic phase. For
shorter simulation runs (right) and therefore reduced accuracy of measurements
the system becomes unstable mocking a two-sublattice orbital ordering pattern.
In addition to the magnetic phases we determined the metallic and insulating
101
6.3 5-band Hubbard model on the Bethe lattice
behavior by investigating the slope of the Green’s function. In the insulat-
ing case the Green’s function has a steep slope and decays exponentially
to G(τ) → 0 at intermediate τ -values, whereas in the metallic case G(β/2)
remains finite and no steep slope is observed.
A Mott-insulating state exists at half-filling as well at n = 4 for J = U/6, so
essentially the Mott insulating state is accompanied by an antiferromagnetic
phase, whereas the metallic phase is either paramagnetic or ferromagnetic.
Since orbital degeneracy stabilizes the metallic phase [94] it becomes plausible
that the ferromagnetic and metallic phase is enlarged compared to the three-
orbital case.
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34
2
3
µ
J
n = 1 n = 2 n = 3 n = 4 n = 5
PM FP FP MI,AFM
Figure 6.7: Phase diagram of the 5-band Hubbard model on the Bethe lattice.
We find a paramagnetic (PM) (green), a ferromagnetic (blue) with fully polarized
parts (FP) (intermediate and dark blue) and an antiferromagnetic (AFM) phase
(orange), where the system is also insulating (MI).
In Figure 6.7 we have summarized the observed phases and sketched the
phase diagram of the 5-orbital Hubbard model on the Bethe lattice. Shaded
green is the paramagnetic phase, orange the antiferromagnetic and Mott-
insulating phase, the remaining part is ferromagnetic. In blue we highlighted
the fully-polarized phase within the ferromagnetic phase and have used a
lighter blue, where the fully-polarized phase starts to develop.
6.3.3.2 Exchange mechanism
When comparing the two different Hund’s coupling strengths J = U/6 and
J = U/4 we see a striking difference at n = 4. Whereas for the larger Hund’s
102
Dynamical Mean Field Theory
coupling J = U/4 the system is ferromagnetic we observe antiferromagnetism
for J = U/6. In order to understand the origin of the different magnetic
phases we looked at the valence histogram.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 1 2 3 4 5 6 7 8 9 10
Pro
bab
ilit
y
Occupation
J=U/6
d4 p2 d4
Figure 6.8: Top: Valence histogram for J = U/6 at n = 4. With probability
Pd8 ∼ 0.91 the system is found in a d4 configuration. Bottom: For systems with
fixed charge occupancy antiferromagnetism due to super-exchange is favorable ac-
cording to the Goodenough-Kanamori-Anderson rules [95]. The antiferromagnetic
ordering is also confirmed by our simulation.
For J = U/6 at density n = 4 the system is most dominantly in a d4 con-
figuration and there are only little contributions from the n = 3 and n = 5
sector as shown in Figure 6.8. Thinking in terms of the lattice system and
considering two sites which are in reach via hopping processes and connected
with each other via some mediator site there will be two 5-orbital impuri-
ties on these sites, which have most likely (with probability P 2d8 ∼ 0.8) the
same valence, namely ntot = 4. The spins on the impurity will arrange such
that the total spin is maximal and at the same time such that the two im-
purities are able to exchange electrons. Concerning the kind of the effective
exchange coupling Goodenough, Kanamori and Anderson [96–98] established
rules (GKA rules) to determine these from the electronic configuration. At
n = 4 the exchange between different sites will be mostly exchange between
103
6.3 5-band Hubbard model on the Bethe lattice
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 1 2 3 4 5 6 7 8 9 10
Pro
bab
ilit
y
Occupation
J=U/4
d4 p2 d3 d3 p2 d4
Figure 6.9: Top: Valence histogram for J = U/4 at n = 4. The system ex-
hibits mixed-valence character. Bottom: For atoms of different valence the double
exchange mechanism leads to an effective ferromagnetic coupling [95]. The ferro-
magnetic ordering is also confirmed by our simulation.
half-filled and half-filled orbitals, which according to the GKA rules leads to
a strong and antiferromagnetic exchange coupling. In Figure 6.8 we depicted
this super-exchange mechanism for a transition metal oxide. Only two or-
bitals of the five orbitals are drawn but the spin orientation of the left-out
three orbitals is shown. The transition metal atoms exchange electrons via
an intermediate oxygen 2p orbital. The favorable antiferromagnetic ordering
can be easily read off from the sketch.
For J = U/4 the valence histogram for n = 4 is qualitatively very different
as shown in Figure 6.9. Still most contributions come from d4, but it is also
quite likely to find the system in a d3 or d5 configuration. Relating again back
to the lattice system this implies that we will have atoms of different valence.
In Figure 6.9 we sketch this exchange mechanism for a doped transition metal
oxide. The underlying exchange mechanism is double exchange and leads to
an effective ferromagnetic exchange interaction as can be seen from Figure
6.9.
104
Dynamical Mean Field Theory
These two examples demonstrate that based on the measured valence his-
togram we can reason the exchange processes obeying Pauli’s exclusion prin-
ciple as well fulfilling Hund’s rule and thereby we can explain the resulting
nature of the magnetic exchange coupling.
6.3.4 Conclusion
In this study we have investigated the phase diagram of the 5-band Hubbard
model on the Bethe lattice for two Hund’s coupling values J = U/4 and
J = U/6 and determined the magnetic as well as metal-insulator behavior
for the range of densities n ∈ [0, 5]. In contrast to a study of the 3-band
version of this model we do not find an orbitally ordered phase, but find a
fully-polarized ferromagnetic phase mocking an orbitally ordered phase when
the accuracy of measurements is reduced.
Depending on the strength of the Hund’s coupling J we have found different
magnetic phases at the density n = 4. We could amount this difference
to the different exchange mechanisms between the atoms and have shown
how starting from the valence statistics we can reason the nature of the
magnetic exchange coupling. As the 5-band Hubbard model on the Bethe
lattice serves as a generic model of transition metal oxides we have thereby
obtained a qualitative picture of various phases which can potentially occur
in these materials.
105
6.4 Extension to real materials: LDA+DMFT
6.4 Extension to real materials: LDA+DMFT
We have argued in Chapter 2 that in order to describe strongly correlated
systems we have to treat atomic-like and band-like behavior on an equal
footing, since these systems are at the brink of localization and delocaliza-
tion. This requirement is met by the Hubbard model which can be solved
approximately using the DMFA.
In order to do first-principles simulations we have to bring in material-specific
properties to this model Hamiltonian. This is achieved by combining the
framework with DFT in the Local Density Approximation and goes under
the name ”LDA+DMFT”. The idea of this approach is to use the band
structure given through the Kohn-Sham eigenvalues for the dispersion ǫ(k)
in Eq. (6.21). What we essentially do is to put the local Coulomb interactions
on top of the non-interacting DFT-solution for the localized orbitals.
6.4.1 The double-counting problem
There is however a problem with this modular building block framework.
Since there are already some correlations considered in LDA via the exchange-
correlation potential some of the interaction contribution are counted twice
and we therefore need to subtract a double counting correction term Edc from
the Hamiltonian. Within LDA the electron density is not given by the density
of the localized orbitals but the total density, which is the reason why this
double counting correction term is not known exactly. We use the Hartree
ansatz for the double counting correction but there are different ansatzes to
compute EDC [99, 100].
For a given compound we usually have a rough estimate of the average oc-
cupation nc in the d-orbitals from its electronic configuration.
The Coulomb energy ECoulomb of these orbitals is given by:
ECoulomb =U
2nc(nc − 1) (6.26)
where we use the screened Coulomb repulsion U . We subtract this expression
from the impurity energy level given by LDA
ǫc =d
dnc(ELDA − ECoulomb) (6.27)
106
Dynamical Mean Field Theory
which gives rise to a double counting shift
EDC = U(nc −1
2). (6.28)
In addition, one also has to subtract the energy contribution coming from
the Hund’s coupling J . As the precise amount is not known, we will scan
possible values close to the estimate of Eq. (6.28) as presented in Section
6.5.
6.4.2 Self-consistency equations
With the newly introduced expressions ǫKSk and EDC for the ”LDA+DMFT”-
framework the k-summation of Eq. (6.21) is adapted to
G(iωn) =∑
k
(iωn + µ− ǫKSk − Σimp(iωn)− EDC
)−1(6.29)
The Kohn-Sham equations form a matrix in orbital space including more
than the correlated (d- or f -) bands, whereas the self-energy and double
counting corrections have non-zero entries only in the blocks corresponding
to the correlated orbitals.
6.4.3 High-frequency expansion
For realistic band structures ǫKSk or lattice DOS other than the semi-circular
DOS, for which the simple Eq. (6.25) holds we have to carry out a Fourier
transform of G(τ) first. In order to compute the Fourier transform we want
to make use of the high-frequency behavior of the Green’s function and hy-
bridization function.
Given a function F (iω) =∫F (τ)eiωnτdω we find by integration by part:
F (iω) = −F (β) + F (0)
iω+F ′(β) + F ′(0)
(iω)2− F ′′(β) + F ′′(0)
(iω)3−
∫F ′′′(τ)
(iω)4eiωτ
(6.30)
At high frequencies the first three terms will be dominant and we write the
high-frequency asymptotic expansion as:
F asym(iω) + F rest(iω) =c1iω
+c2
(iω)2+
c3(iω)3
+O( 1
(iω)4) (6.31)
with the coefficients ci identified with the numerators of the terms in Eq.
(6.30).
107
6.4 Extension to real materials: LDA+DMFT
If we Fourier transform F (iω) to the imaginary time we obtain:
F (τ) = −c12+c24(−β − 2τ) +
c34(βτ − τ 2) + Frest(τ) (6.32)
The Fourier transform of the Green’s function G(τ) and the hybridization
function ∆(τ) enter the self-consistency equations. What we want to do is
subtract the asymptotic part F asym(τ) from the input function in imaginary
time F (τ) and numerically compute the Fourier transform of Frest(τ). To
F rest(iω) we again add the high-frequency behavior F asym(iω) in order to
obtain F (iω). The advantage is that the numerical Fourier transform is
much more well-behaved for this remaining term.
One possibility is to use the numerical derivatives to compute the coefficients,
which may however be afflicted with a large error and we will use the analytic
expressions for c1 and c2 of the Green’s function.
Given the definition of the Green’s function
G(τ) = −〈Tτ [d(τ)d†(0)]〉 (6.33)
with d(†) given in the interaction representation d(τ) = eHτdeHτ we can
formally expand the exponential in a Taylor series:(
1 +Hτ +H2τ 2
2!+ ...
)
d
(
1−Hτ + H2τ 2
2!+ ...
)
=
d+ τ [H, d] +τ 2
2![H, [H, d]] + ...
(6.34)
If we insert this expression in the definition of the Green’s function in Eq.
From the fermionic anti-commutation rule, we immediately find c1 = 1.
For the higher moments we have to be explicit on the Hamiltonian we choose.
We consider a Hamiltonian containing the four-fermion Coulomb interaction
matrix.
H =∑
σ
ǫdσd†σdσ+
∑
k,σ
ǫk,σc†kσckσ+
∑
k,σ
(Vkσckσdσ + h.c)+∑
i,j,k,l
Uijkld†iσd
†jσ′dlσ′dkσ
(6.36)
108
Dynamical Mean Field Theory
which we investigated in a study of transition metal monoxides described in
Section 6.5. We will first consider those terms in the Hamiltonian with two
fermionic operators. The bath and hybridization terms do not contribute
since [c†kσckσ, dσ′] = 0 and [ckσdσ, dσ′ ], d†σ′ = 0. Only the part describing
the impurity delivers
ǫdσ[d†σdσ, dσ′
]= ǫdσdσ. (6.37)
For the evaluation 〈[HU , d], d†〉 of the four-fermion part
HU =∑
i,j,k,lUijkld†iσd
†jσ′dlσ′dkσ it is useful to expand the expression in an-
ticommutators of the form
d(†)α , d
(†)β
which we can immediately evaluate
using the fermionic anticommutation rules. The following identities are use-
ful to determine the contribution of HU to c2: For the commutator [H, d] we
use
[ABCD,E] = AB [CD,E] + [AB,E]CD
[AB,C] = A B,C+ A,CB(6.38)
and
ABC,D = AB C,D − A B,DC + A,DBC (6.39)
for the anti-commutator 〈[H, d], d†〉. Carrying out the commutator expan-
sion and evaluation and combining the result with Eq. (6.37) we obtain:
c2 = −ǫd −∑
j
(Uijij − δσσ′Uijji) 〈ninj〉 (6.40)
with
〈ninj〉 = 〈d†iσd†jσ′djσ′diσ〉 (6.41)
109
6.5 Transition metal monoxides: NiO and CoO
6.5 Transition metal monoxides: NiO and CoO
6.5.1 Introduction
To the presented study the following persons have contributed: B. Surer, A.
Poteryaev, Jan Kunes, M. Troyer, A. Lauchli and P. Werner.
Transition metal oxides are a class of materials which has been investigated
extensively, both experimentally and theoretically. These compounds exhibit
a multitude of phases with remarkable properties, including high tempera-
ture superconductivity, colossal magnetoresistance and Mott metal-insulator
transitions [101]. Central to the physics of TMOs is the question of electronic
correlations and their interplay with chemical bonding. The late transition
metal monoxides are prototypical systems for the investigation of this is-
sue. Classified as charge-transfer insulators in the Zaanen-Sawatzky-Allen
scheme [102], they are characterized by the strong influence of the hybridiza-
tion between the transition metal 3d and oxygen 2p orbitals on their physical
properties. In this study we will consider two prominent charge-transfer in-
sulators, NiO and CoO.
The basic understanding of the electronic spectra of these materials dates
back to the cluster ED studies of Fujimori et al. [103] and Sawatzky and
Allen [104]. The first material specific calculations based on ab-initio DFT
were those of Anisimov, Zaanen and Andersen [105] using the so called
LDA+U method. Compared to pure DFT calculations, LDA+U signifi-
cantly improved the description of physical properties related to the elec-
tronic ground state and their adiabatic changes such as equilibrium lattice
constants, bulk moduli or phonon spectra, but this description fails to cap-
ture some qualitative features in the electron excitation spectra. Recently,
the application of many-body techniques in the framework of DMFT lead
to substantial improvement of the theoretical description of charge-transfer
systems.
The DMFT results obtained so far still involve several approximations, whose
effect is not well understood. In this study we address one of them, namely,
the form of the on-site interaction. This is a particularly important ques-
tion since it involves symmetry. We use the Krylov implementation of the
Hybridization expansion solver with full Coulomb interaction (having the
SU(2) symmetry) to perform LDA+DMFT calculations on the paramag-
netic phases of NiO and CoO. The results are compared with calculations in
110
Dynamical Mean Field Theory
the density-density approximation, calculations performed with ED as well
to Photo-emission spectra of the two systems.
6.5.2 Model and Computational method
In order to account for the charge-transfer character of the studied transition
metal monoxides we investigate an eight-band p-d Hamiltonian defined in
Eq. (2.20). For the construction of the effective Hamiltonian, for which LDA
provides the band structure, the projection onto Wannier functions has been
applied [106].
We investigate the systems CoO and NiO for three different formulations of
the Coulomb interaction. These three cases include the density-density case
of Eq. (2.18), the Slater-Kanamori case of Eq. (2.17) and the full Coulomb
interaction case of Eq. (2.7).
We solve Eq. (2.20) within the single-site DMFA using the Krylov implemen-
tation of the hybridization expansion to solve the quantum impurity model.
As shown in previous Sections this method is capable of treating a five orbital
model with general four-fermion interaction terms in the strong correlation
regime. We examined the effect of truncating the outer trace by comparing
the results of a single DMFT iteration for different values of Ntr as discussed
in Section 3.3.3.2. We considered states up to an energy +2 eV above the lo-
cal ground state energy and confirmed by comparing the observables that it is
sufficient to perform the trace over the states of the degenerate ground-state.
The simulations were performed at a temperature β−1 = 0.1 eV and interac-
tion values parametrized by U = 8 eV and J = 1 eV for NiO and U = 8.8
eV and J = 0.92 eV for CoO. The Green’s function was accumulated on 200
time slices on the imaginary time axis, which is appropriate at this temper-
ature. Since the electronic configuration of NiO is 2s22p63d8 and 2s22p63d7
for CoO, we expect a total number of 14 (NiO) and 13 (CoO) electrons in
the O 2p and Ni 3d orbitals. In the DMFT calculation, the chemical poten-
tial is adjusted such that the self-consistent total density corresponds to this
expected total number of electrons.
For the double counting value we have scanned possible values around the
estimate of Eq. (6.28) yielding EDC ∼ 60 eV (NiO) and EDC ∼ 57 eV (CoO)
and chose the values EDC = 59 eV (NiO) and EDC = 57 eV (CoO). Higher
double counting values yield a metallic system.
NiO and CoO have a rock-salt crystal structure. Due to the presence of the
111
6.5 Transition metal monoxides: NiO and CoO
oxygens the degeneracy of the 3d orbitals is lifted and the levels are split into
two eg and three t2g orbitals. LDA yields for the crystal field splittings the
values ∆cf = 0.566 eV (NiO) and ∆cf = 0.273 eV (CoO).
6.5.3 Results
6.5.3.1 Occupation
With the chosen double counting we find the hybridized Ni 3d - O 2p (Co 3d
- O2p) orbitals to have an occupation n = 8.21(1) (n = 7.14(1)) where the t2gorbitals have an average occupation of nt2g = 0.99982(7) (nt2g = 0.8180(8))
and neg = 0.553(2) (neg = 0.558(2)) for the eg orbitals. In Figure 6.10
we have depicted the electron configuration of NiO and CoO. According to
Hund’s rule the d8-configuration (NiO) is expected to have spin S = 1 and
the d7-configuration (CoO) spin S = 3/2.
eg
t2g∆cf
eg
t2g∆cf
Figure 6.10: Diagram of the electron configuration of NiO (left) and CoO (right).
6.5.3.2 Sector and State statistics
The degeneracy of the spin triplet state in the spin-rotation symmetry pre-
serving simulation can be seen in the left panel of Figure 6.11 presenting the
sector statistics. The sector statistics provides the statistics about the proba-
bility of the impurity to be in different quantum states. For the analysis, it is
most convenient to group the different eigenstates according to the conserved
quantum numbers (nup, ndown), n = 0, . . . , 5, resulting in a histogram repre-
senting the weight in 36 quantum number sectors. The spin triplet ground
state has 2 holes in the eg orbital for spin up (↓↓〉), 2 holes for spin down
(↑↑〉) and one hole in the eg orbital for both spin directions (↑↓ + ↓↑)/√2.
In the density-density case the spin rotation symmetry is broken and the
ground state is only doubly degenerate, because the (↑↓ + ↓↑)/√2 state lies
higher in energy. The preserved degeneracy of the spin triplet ground state
is reflected in the fact that the three sectors (3, 5), (4, 4), (5, 3) belonging to
the d8 configuration dominate the histogram with three equal contributions.
112
Dynamical Mean Field Theory
0
0.05
0.1
0.15
0.2
0.25
0.3
0 5 10 15 20 25 30 35
Pro
babilit
y
Sector
sector statistics
0
0.05
0.1
0.15
0.2
0.25
0.3
0 5 10 15 20 25 30 35
Pro
babilit
y
Sector
sector statistics
0
0.02
0.04
0.06
0.08
0.1
0 1 2 3 4 5
Pro
bab
ilit
y
State
Sector (2,5)Sector (3,4)Sector (4,3)Sector (5,2)
Figure 6.11: Sector statistics of NiO (left) and CoO (middle) in the full Coulomb
interaction formulation. Right: Atomic weights of the states within the 7-particle
sectors of CoO. The weights of the states above the 12-fold degenerate ground-state
are negligible. Only the 5 lowest states are shown.
The spin rotation symmetry is also preserved in the full Coulomb matrix
simulation of CoO and seen in the degeneracy of the atomic weights, displayed
in the middle panel of Figure 6.11.The degeneracy of the dominant sectors
(2, 5), (3, 4), (4, 3), (5, 2) of the d7 configuration are well in line with a spin
quartet S = 3/2 state. Applying Hund’s rules to the electron configuration
of CoO we expect a hole in one of the three t2g levels of the Co 3d orbitals
giving rise to a three-fold degeneracy for each of the 4 sectors, see sketch in
Figure 6.10. Indeed, as can be seen in the right panel of Figure 6.11, which
plots the weights of individual states in the 7-particle sectors, the resulting
12-fold degeneracy of the ground-state is confirmed by the degeneracy of the
atomic weights.
6.5.3.3 Density of states
To obtain the single-particle spectral functions analytic continuation to real
frequencies is performed using MaxEnt as explained in Chapter 4. The spec-
tral functions of NiO within the density-density, Slater-Kanamori and full
Coulomb interaction formulation are shown in Figure 6.12. There is hardly
any difference in the spectral function computed within the density-density
approximation and the two SU(2) invariant formulations of the Slater-Kanamori
and the full Coulomb interaction. In order to exclude the possibility that
the conformity of spectral functions is due to any smearing out of spectral
features inherent to MaxEnt we have made a comparison of these three for-
mulations using ED as an impurity solver. The advantage is that ED works
directly with real frequencies and does not require analytic continuation of
the data. Unlike CTQMC it can incorporate only a few bath sites, hence
113
6.5 Transition metal monoxides: NiO and CoO
0
0.1
0.2
0.3
0.4
0.5
−20 −15 −10 −5 0 5 10 15 20
Spec
tral
den
sity
Energy (eV)
nn
egt2g
total
0
0.1
0.2
0.3
0.4
0.5
−20 −15 −10 −5 0 5 10 15 20
Spec
tral
den
sity
Energy (eV)
SK
egt2g
total
0
0.1
0.2
0.3
0.4
0.5
−20 −15 −10 −5 0 5 10 15 20
Spec
tral
den
sity
Energy (eV)
full U
egt2g
total
0
0.1
0.2
0.3
0.4
0.5
−20 −15 −10 −5 0 5 10 15 20
Spec
tral
den
sity
Energy (eV)
full UnnSK
Figure 6.12: Spectral function of NiO in the density-density (nn) approximation,
with the Coulomb interaction in the Slater-Kanamori (SK) form and with the full
four-fermion (full U) Coulomb interaction.
effects of the hybridization will not be adequately captured. However, con-
cerning differences in the local interaction we expect ED to detect the dif-
ferences very clearly. As a matter of fact we expect it to rather overestimate
these differences because the on-site part is fully incorporated whereas the
bath and hybridization is only partially considered.
The comparison of different Coulomb interaction formulation computed within
ED is shown in Figure 6.13. Again, we encounter qualitatively very similar
spectral functions for the three formulations. Detecting the tiny differences
of the presented spectral functions also within CTQMC is very unlikely due
to the effect of the bath as well due to analytic continuation. In hindsight it
is therefore not surprising that all three formulations yield qualitatively the
same result in the CTQMC-calculations.
114
Dynamical Mean Field Theory
0
0.2
0.4
0.6
0.8
1
−16 −14 −12 −10 −8 −6 −4 −2 0 2 4 6
DO
S
E (eV)
Full USlater KanamoriDensity−Density
0
0.2
0.4
0.6
0.8
1
1.2
−16 −14 −12 −10 −8 −6 −4 −2 0 2 4 6
DO
S
E (eV)
CTQMCED
PES
Figure 6.13: Left: Spectral function of NiO obtained from ED for all three
formulations of the interaction. Right: ED and CTQMC Spectral function of NiO
compared to PES [104].
We compare the computed spectral function of NiO to the PES data of
Sawatzky et al. [104]. Eastman and Freeouf studied the 2p contribution of
the oxygen in the spectra of NiO and showed that it decreases with increasing
photon energy [107]. At 120 eV photon frequency the O 2p contribution is
negligible at which frequency the PES-data by Sawatzky et al. has been ob-
tained. The comparison between ED, CTQMC and PES is shown in Figure
6.13. First of all the agreement between the ED and the analytically contin-
ued CTQMC data is good except for the double peak structure at around
−2 eV and −4 eV. As we have already noted for the spectral function of Fe
on Na presented in Section 5.2 the analytically continued CTQMC data fails
to resolve a similar double peak structure at these frequencies. When com-
paring the computed spectral function to the PES data we see that both ED
and CTQMC underestimate the experimentally found gap of NiO of the or-
der of ∆ ∼ 4 eV [104,108] which suggests that the estimate for the Coulomb
interaction U is below the appropriate value. At energies below the Fermi-
level (EF = 0) the theoretical curves agree quite well with the experimental
data. The spectral features at −2 eV and the broad part at around −8 eVto −10 eV are reproduced by the QMC and ED data, where ED in addition
captures the feature at −4 eV.In Figure 6.14 the spectral functions of CoO within the full Coulomb inter-
action formulation is shown. The spectral function has a similar shape as the
NiO data with the difference that the incoherent part extends to frequencies
115
6.5 Transition metal monoxides: NiO and CoO
0
0.1
0.2
0.3
0.4
0.5
−20 −15 −10 −5 0 5 10 15 20
Spec
tral
den
sity
Energy (eV)
egt2g
total
0
0.2
0.4
0.6
0.8
1
−16 −14 −12 −10 −8 −6 −4 −2 0 2 4 6
DO
S
E (eV)
CTQMCLDA O p
PES
Figure 6.14: Left: Spectral function of CoO with the full Coulomb interaction.
Right: Spectral function of CoO compared to PES [108].
down to −15 eV. We compare our data to the experimental data of Parmi-
giani [108] shown in Figure 6.14. Also the incoherent part of the PES-data
has weight at these large frequencies. The computed spectral function how-
ever looks qualitatively very different from the PES-spectra presented in the
right panel of Figure 6.14. Ref. [108] did not provide information on how
much spectral weight in the spectra comes from the oxygen. From the LDA
calculation we see that the energy range where the O 2p contributions would
enter is exactly where the CTQMC data lacks the spectral weight compared
to the PES-data. We find the gap to be of the order ∆ ∼ 2.8 eV which is in
good agreement with the experimental value of ∆ ∼ 2.6 eV [108].
6.5.4 Conclusion
We showed that the ground-state degeneracy of both NiO and CoO is cor-
rectly captured by our spin-rotation-symmtery preserving CTQMC-simulation.
For the spectral function of NiO within the different formulations of the
Coulomb interaction we find neither in CTQMC nor in ED any appreciable
difference. Pronounced features in the PES data are well reproduced by the
computed spectral functions, where mismatches can be probably related to
the additional weight coming from the oxygens. We find that simulations of
charge-transfer systems constitute an ambitious goal, because it is difficult
to capture physical effects coming from both the Ni 3d (Co 3d) and the O 2p
orbitals in a correct balance as well to disentangle these effects in the PES
116
Dynamical Mean Field Theory
data, which we rely on for our comparison. These systems therefore seem
not to be an optimal starting point for first-principles simulations as there
are too many adjustable parameters such as the double counting value in
addition to the material specific input parameters.
117
6.5 Transition metal monoxides: NiO and CoO
118
Chapter 7
Conclusion
In this thesis we have studied transition metal compounds from first-principles
taking into account the full intra-atomic Coulomb interactions of these 5-
orbital systems. As the physics of these transition metal compounds is de-
termined by strong electron correlations we have treated these systems in a
many-body formulation. We have explored the first-principles capabilities of
“DFT+Model Hamiltonian” descriptions of quantum impurity systems and
applied the recently developed Krylov implementation of the hybridization
expansion algorithm to solve the quantum impurity models numerically ex-
act. We have addressed both lattice systems, i.e., solids which we treated
within the DMFA as well as genuine quantum impurity systems.
Concerning genuine quantum impurity systems we have studied Iron impu-
rities on Sodium (Fe on Na) as well as Cobalt impurities in a Cupper host
(Co in and on Cu ).
For the Fe on Na system we have identified the transitions in the spectral
function and have shown how the reservoir of conduction electrons of the
Na host acts to screen additional charges on the Fe impurity. If we finetune
the occupation on the Fe impurity, we found our CTQMC calculations to
reasonably reproduce the PES-spectra and conclude that improvements to
better predict the chemical potential are needed. In this study we have
found ED and CTQMC a synergetic combination to study spectral functions,
because it enabled us to consider a continuous bath and to still identify the
underlying transitions in the spectra, as well as to judge the quality of the
analytical continuation of the CTQMC data.
In a study of Co in and on Cu we investigated the multi-orbital Kondo
effect and computed the Kondo temperature from first principles. We treated
119
these systems as multi-orbital Anderson impurity models with the full four-
fermion Coulomb interaction and observed that their physical behavior is
beyond the physics of simple “spin-only” models. In our study we have
shown that the Kondo temperature, TK , associated with the onset of Fermi
liquid behavior can differ between the impurity orbitals. The comparison
of our QMC calculations and scaling arguments further demonstrated that
fluctuations in the orbital degree of freedom and Hund’s rule coupling are
crucial in determining TK in realistic systems. We believe that the role of
orbital degrees of freedom is also of importance for the understanding of
other magnetic nanostructures based on 3d adatoms on surfaces as well as
magnetic impurities in bulk metals.
Concerning strongly correlated materials we have focussed on transition metal
oxides and studied a generic model of transition metal oxides (five band Hub-
bard model on the Bethe lattice) as well as the transition metal monoxides
Nickeloxide (NiO) and Cobaltoxide (CoO).
In the study of the 5-band Hubbard model on the Bethe lattice we have
determined the magnetic as well as the metal-insulator behavior for different
densities and compared two cases for the Hund’s coupling J . We could
amount differences in the magnetic phases to different exchange mechanisms,
which also occur in transition metal oxides. We have shown how starting from
the computed valence statistics we can reason the nature of the magnetic
exchange coupling and conclude that the 5-band Hubbard model on the Bethe
lattice is potentially very interesting to reproduce and provide insights to
aspects typical of transition metal oxides.
Concerning our study of NiO and CoO we have demonstrated that the
ground-state degeneracy of both NiO and CoO is correctly captured by our
spin-rotation-symmtery preserving CTQMC-simulation. We have found that
pronounced features in the PES data are well reproduced by the computed
spectral functions and motivated why mismatches to the experimental data
can probably related to the additional weight coming from the oxygens.
Concerning the aim to establish first-principles simulations we suggest to
further explore quantum impurity systems whose studies are by far not yet
exhausted. They provide an optimal starting point for first-principles simu-
lations as they have only few parameters to be determined ab-initio. We
believe that detailed benchmarks and the resulting improvements to the
“DFT+Model Hamiltonian” approach represent an indispensable step to-
wards achieving first-principles simulations of strongly correlated materials.
120
Bibliography
[1] E. Schrodinger, “Quantisierung als Eigenwertproblem,”
Annalen der Physik 385, 437–490 (1926)
[2] P. Hohenberg and W. Kohn, “Inhomogeneous Electron Gas,”
Phys. Rev. 136, B864–B871 (Nov 1964)
[3] W. Kohn and L. J. Sham, “Self-Consistent Equations Including Ex-
change and Correlation Effects,” Phys. Rev. 140, A1133–A1138 (Nov
1965)
[4] Felix Bloch, “Uber die Quantenmechanik der Elektronen in Kristallgit-
tern,” Zeitschrift fur Physik A Hadrons and Nuclei 52, 555–600 (1929)
[5] A. Sommerfeld, “Zur Elektronentheorie der Metalle auf Grund der Fer-
mischen Statistik,” Zeitschrift fur Physik A Hadrons and Nuclei 47,
1–32 (1928)
[6] H. Bethe, “Theorie der Beugung von Elektronen an Kristallen,”
Annalen der Physik 392, 55–129 (1928)
[7] J H de Boer and E J W Verwey, “Semi-conductors with
partially and with completely filled 3 d -lattice bands,”
Proceedings of the Physical Society 49, 59 (Jan 1937)
[8] N F Mott and R Peierls, “Discussion of the paper by de Boer and
Verwey,” Proceedings of the Physical Society 49, 72 (Jan 1937)
[9] G. Kotliar, S. Y. Savrasov, K. Haule, V. S. Oudovenko, O. Parcollet,
and C. A. Marianetti, “Electronic structure calculations with dynami-
cal mean-field theory,” Rev. Mod. Phys. 78, 865–951 (Aug 2006)