3 The Physics of Quantum Impurity Models

Jan von Delft

Arnold Sommerfeld Center for Theoretical Physics
Ludwig-Maximilians-Universität München
Theresienstr. 37, 80333, Munich

Contents
1 Introduction 2

2 Single-impurity Anderson model: local moment formation 3

3 Kondo model: spin-exchange interaction 5

4 Numerical renormalization group 8

5 NRG results for the single-impurity Anderson model 12

6 Two-channel Kondo model 17

A Schrieffer-Wolff transformation 19

E. Pavarini, E. Koch, A. Lichtenstein, and D. Vollhardt (eds.)

Dynamical Mean-Field Theory of Correlated Electrons
Modeling and Simulation Vol. 12
Forschungszentrum Jülich, 2022, ISBN 978-3-95806-619-9
http://www.cond-mat.de/events/correl22
3.2 Jan von Delft

1 Introduction
A quantum impurity model describes a localized, discrete quantum system, the impurity, cou-
pled to noninteracting excitations with a continuous (non-discrete) excitation spectrum, the
bath. Such models were first invoked in the 1960s to explain certain anomalies in the resis-
tivity of magnetic alloys, where the impurities are localized magnetic moments (or localized
spins, for short). In subsequent decades, quantum impurity models turned out to be relevant in
several other contexts, including transport through quantum dots and nanotubes, whose discrete
energy levels act as impurities; dissipation and decoherence of qubits, described as impurities
coupled to bosonic baths; and dynamical mean field theory, where a single lattice site is viewed
as an impurity coupled to a self-consistently determined bath. This lecture aims to introduce
the physics of quantum impurity models from a renormalization group perspective. By way of
introduction, we begin with some historical remarks.
In 1934, de Haas, de Boer and van den Berg [1] discovered an unexpected phenomenon when
measuring the resistance of gold as a function of temperature, ρ(T ): it showed a minimum at
T = 3.70 K, whereas it had been expected that the resistance would decrease monotonically
with decreasing temperature as phonons freeze out, reaching its smallest value at T = 0. The
fact that ρ(T ) instead turns upward once T drops below 3.70 K indicates that in this regime
the electron scattering rate increases with decreasing temperature—a finding that was very sur-
prising at the time. In subsequent years, similar resistance minima were found in numerous
so-called dilute magnetic alloys: metals such as Cu, Ag, Au containing a small concentration
of magnetic impurities such as Cr, Mn, Fe.
The origin of the resistance minima observed for magnetic alloys remained puzzling until 1964,
when Kondo offered an explanation [2]. He considered a model, since known as the Kondo
model, describing a dilute concentration of localized spin-1/2 impurities immersed in a metallic
conduction band, with a spin-exchange coupling between the localized spins and the conduction
band. This coupling causes spin-flip scattering events: when a conduction electron scatters off
a localized spin, the spin of both can flip. Kondo computed the corresponding scattering rate
to lowest nontrivial order in the exchange coupling, and found that it increases with decreasing
temperature, causing the resistance ρ(T ) to likewise increase. This explained the resistance
minimum.
Moreover, Kondo found that the scattering rate, γ(T ), increased logarithmically with decreasing
T , like ln(D/T ), where D is a large energy scale such as the bandwidth of the conduction
band. This meant that while he had resolved an experimental puzzle, he had simultaneously
discovered a theoretical one, which came to be known as the Kondo problem: what happens
to the scattering rate in the limit T → 0? The perturbative result, implying a logarithmic
divergence, clearly can not be trusted in this limit.
To answer this question, nonperturbative methods were called for. Over the years, a wide variety
of approaches were applied to the Kondo problem, including the numerical renormalization
group (NRG) [3, 4], the Bethe Ansatz [5, 6], conformal field theory [7], bosonization [8, 9],
and Monte Carlo methods [10]. As a result, the answer to the above question is now very well
Impurity Models 3.3

established: the scattering rate γ(T ) approaches a constant value for T → 0, because the local
spin is screened by conduction electrons to form a spin singlet. The screened singlet acts just
like a static (i.e. nonmagnetic) impurity, off which other electrons scatter very strongly, albeit
without flipping their spin. The screening of a localized spin by conduction electrons has come
to be known as the Kondo effect.
The crossover, with decreasing T, from an unscreened spin to a spin singlet is a paradigmatic
example of a renormalization group (RG) flow between two RG fixed points. The goal of these
notes is to elucidate the physics of quantum impurity models from such an RG perspective.
In the following sections, we will introduce two paradigmatic models describing spin screening:
the single-impurity Anderson model and the Kondo model; explain why a perturbative treat-
ment fails at low temperatures; describe Wilson’s numerical renormalization group approach
for reaching low temperature; and discuss various physical quantities exhibiting signatures of
spin screening. In the final section, we briefly discuss a model that does not show full spin
screening—the two-channel Kondo model. We set ~ = 1 and kB = 1 throughout.

2 Single-impurity Anderson model: local moment formation

In 1961, P.W. Anderson introduced a “highly simplified” model, since known as the single-
impurity Anderson model (SIAM), to explain the formation of local moments in magnetic al-
loys. In such systems, the magnetic impurities have d or f orbitals that are strongly localized,
yet nevertheless hybridize with the host metal, in such a manner that they behave as localized
spins (or magnetic moments) acting as dynamical degrees of freedom. A fully realistic descrip-
tion is very challenging, since d or f shells typically host several electrons, with strong local
Coulomb interactions. Anderson argued that the essential physics could be captured by con-
sidering just a single, local level rather than an entire shell. He thus studied the Hamiltonian
HSIAM = Hbath + Hloc + Hhyb , with

εd + U ∆

vk
εF = 0
vk εk
εd ∆

X
Hbath = εk n̂ks , n̂ks = c†ks cks , (1a)
ks
X

Hloc = εd − 12 hs n̂ds + U n̂d↑ n̂d↓ , n̂ds = d†s ds , (1b)
s
X

Hhyb = vk c†ks ds + d†s cks . (1c)
ks
3.4 Jan von Delft

Here, Hbath describes a band of free (noninteracting) electrons with spin s ∈ {↑, ↓} and energy
εk (k is a continuous momentum index), Hloc describes a discrete, localized level with energy εd ,
spin s, Zeeman splitting h if a magnetic field is applied (we set gµB = 1), and Coulomb energy
cost U if the level is doubly occupied. Hhyb describes hybridization between the d level and the
conduction band, allowing electrons to hop back and forth between the two with amplitude vk
(we choose it real and constant, vk = v). As a result, the d level acquires a finite width, ∆.
Since Hbath and Hhyb are quadratic, the conduction band can be integrated out, in principle. One
then finds that the dependence of the d level dynamics on bath parameters enters only via the
hybridization function
X vk2
∆(ω) = . (2)
k
ω − εk + i0+

R Γ ()
When this function is expressed through its spectral representation, ∆(ω) = d ω−+i0 + , the

corresponding hybridization spectrum is given by

X
Γ () = |vk |2 δ(−εk ) . (3)
k

It characterizes how strongly a conduction band level with energy  couples to the impurity.
Anderson assumed the hybridization spectrum to be constant near the Fermi energy; for con-
creteness, we will use the box-shaped form,


Γ () = (∆/π) Θ D−|| , ∆/π = v 2 ν . (4)

Here, D is the half-bandwidth of a flat band centered on the Fermi energy at εF = 0, related
to the density of states per spin, ν, by ν = 1/(2D), and ∆ is the width acquired by the d level
through hybridizing with the bath. (This follows from Eq. (23) below.) Unless stated otherwise,
we measure energy in units of D = 1.
The Hilbert space of the local level is spanned by four states, |0i, |↑i, |↓i, |↑↓i, describing the
local level being empty, occupied by one electron with spin ↑ or ↓, or doubly occupied, respec-
tively. They are eigenstates of Hloc , with eigenenergies E0 = 0, E↑ = d − 21 h, E↓ = d + 12 h,
and E↑↓ = 2εd + U, respectively. Unless stated otherwise, we will consider the case of zero
magnetic field, h = 0. We are interested in the local moment regime, where the parameters are
chosen such that the empty and doubly-occupied levels lie well above the (broadened) singly-
occupied levels; this implies E0 −Es > ∆ and E↑↓ −Es > ∆, i.e.

εd + ∆ < 0, εd + U > ∆, (5)

P
Then, the average occupancy of the local level, nd = s hn̂ds i, is ' 1. Thus, the d level hosts a
local moment, containing a spin-up or down electron with equal probability, nd↑ = nd↓ ' 21 .
Transitions between these two spin states can occur via second-order hopping processes involv-
ing |0i or |↑↓i as virtual, high-energy intermediate states. Their net result is that the spin on the
Impurity Models 3.5

d level has flipped from |↑i to |↓i, or vice versa, and a particle-hole excitation has been created
in the band. The effective spin-flip rate due to such processes is found to be
v2 v2 U v2 2U ∆
J= − =− =− . (6)
εd + U εd εd (εd + U ) πεd (εd + U )
Note that J > 0, since in the local moment regime of Eq. (5), εd < 0. If one is interested only
in the physics of these spin-flip processes, it is convenient to consider the Kondo model, which
we discuss next.

3 Kondo model: spin-exchange interaction

The Kondo model (KM) is obtained from the SIAM by projecting the latter onto the subspace
in which the local occupancy is strictly nd = 1, i.e., only the impurity states |↑i and |↓i are
considered, and focusing on the scattering of low-energy band excitations only. Formally, this
can be achieved using a so-called Schrieffer-Wolff [11] transformation (outlined in App. A).
Then, one obtains an effective Hamiltonian of the form HKM = Hbath + Hspin + Hexchange , with
X
Hbath = εk n̂ks , Hspin = hŜzd , (7a)
ks
X
Hexchange = J Ŝd · ŝc , ŝc = c†ks 21 σss0 ck0 s0 . (7b)
ks,k0 s0

This defines the KM. Here, Hbath again describes a free conduction band; Ŝd are spin- 12 op-
erators acting in the space spanned by |↑i and |↓i; and ŝc are conduction band spin operators,
describing its spin density at the impurity site. The local spin couples to the conduction electron
spin density via an antiferromagnetic exchange interaction, J > 0, favoring an anti-alignment
of the spins of the impurity and the conduction band. The KM describes the low-energy be-
havior of the SIAM, at energy scales for which charge fluctuations involving |0i and |↑↓i are
“frozen out”, i.e., can occur only virtually.
The KM is an interesting model in its own right. It has been studied using a wide variety of
theoretical methods, both analytical and numerical, and can be regarded as thoroughly under-
stood. Below, we briefly summarize some of its salient properties. These will be elaborated in
subsequent sections.
Perturbative treatment: When the KM is used to perturbatively compute the scattering rate of
conduction electrons off the local spin in an expansion in powers of J, this rate is found to
increase logarithmically with decreasing temperature, γ(T ) ∼ J + νJ 2 log(D/T ).
Effective coupling: The perturbation expansion can be expressed through an effective, T -
dependent dimensionless coupling, g(T ), having the form
1
g(T ) = , g0 = νJ . (8)
1/g0 − ln(D/T )
At large temperatures, T ' D, the effective coupling reduces to its bare value, g0 . However, it
increases with decreasing T due to the logarithm in the denominator.
3.6 Jan von Delft

Kondo temperature: When the temperature becomes sufficiently small, the perturbative cor-
rection is no longer small and the perturbative treatment breaks down (becomes invalid). The
characteristic crossover scale at which this happens is called the Kondo temperature, TK . One
way of defining it is as the temperature at which the effective coupling diverges, g(TK ) = ∞,
i.e., where the denominator in Eq. (8) vanishes, 1/g0 = ln(D/TK ). This yields

TK ∼ De−1/(Jν) . (9)

The Kondo temperature is exponentially small in the exchange coupling, reflecting the fact that
g(T ) depends logarithmically on T .
Universality: For T > TK , the effective coupling can be expressed in the form
1 1
g(T ) = = . (10)
ln(D/TK ) − ln(D/T ) ln(T /TK )

(g(T ) is not defined for T < TK , since its perturbative derivation becomes invalid.) This
beautiful expression shows that the physics depends on the bare parameters of the model, J,
ν and D, only in the combination TK . When physical quantities, such as the resistivity or spin
susceptibility are measured for different materials, or computed using different bare parameters,
they will collapse onto a universal scaling curve when plotted as functions of T /TK .
Spin singlet ground state: The regime T < TK is beyond the reach of perturbative methods.
However, the nature of the ground state can be anticipated from the fact that the effective cou-
pling becomes large with decreasing temperature: the ground state will be such that the ex-
change interaction in the Hamiltonian, Ŝd · ŝc , is minimized. Consequently, the ground state of
the KM is a spin singlet, i.e., has total spin S = 0.
Kondo cloud: The singlet ground state can be visualized as follows: A cloud of conduction
electrons with a net spin of 1/2, the so-called Kondo cloud, binds the local spin 1/2 into a
singlet. Commonly used schematic depictions of this singlet include the following:

 
d c , √1 | | −| |  , d ξK .
2

The singlet may be viewed as a bound state with binding energy TK , albeit not a very tightly
bound one: since TK is an exponentially small energy scale, the spatial extent of the Kondo
cloud scales is exponentially large, scaling as ξK ∼ vF /TK , where vF is the Fermi velocity.
Spin screening: The effect of the Kondo cloud on the local spin is called spin screening: by
binding the local spin into a singlet, the cloud hides the presence of a local dynamical degree
of freedom from the rest of the system. This has striking consequences for the temperature
dependence of various physical quantities: as the temperature is decreased from above to below
Impurity Models 3.7

TK , their behavior changes from reflecting the presence of a free spin to reflecting its absence
(since it is screened into a singlet). Correspondingly, T  TK is called the local moment (LM)
regime, and T  TK the strong-coupling (SC) regime.
We next discuss the leading T dependence of three physical quantities in these regimes: the
impurity entropy S(T ), the impurity spin susceptibility χ(T ), and the spin scattering rate, γ(T ).
Corresponding numerical results will be shown in Sec. 5 when discussing the SIAM, whose
low-energy behavior matches that of the KM.
Impurity entropy: The high- and low-temperature limits of the impurity entropy are given by

ln(2) , T  TK ,
S(T ) ' (11)
ln(1) = 0 , T  TK .

These reflect the fact that the free, unscreened impurity has two degenerate states, |↑i and |↓i,
whereas the ground state is a non-degenerate singlet.
Spin susceptibility: The static spin susceptibility, defined as the linear response of the local spin
to an applied magnetic field, has the following limiting behaviors:
 1 

z
d Sd (h) T  1 − O ln(T /T K ) , T  TK ,
χ(T ) = ' 4T   (12)
dh χ(0) 1 − O(T 2 /T 2 ) , T T .
h=0 K K

For high temperatures, it shows the Curie behavior, χ(T ) ∼ 1/T, characteristic of a free spin,
with logarithmic corrections reflecting the onset of spin screening. As spin screening becomes
stronger with decreasing temperature, the Curie behavior is cut off, and for T → 0 the suscep-
tibility approaches a constant, comparable in magnitude to 1/TK . In fact, this constant can be


used to define the Kondo temperature, setting TKχ = 1/ 4χ(0) [4, 12]. (This definition has the
advantage that it does not depend on bare parameters of the model. It differs from that of Eq. (9)
by a constant of order unity, but that is no cause for concern, since the Kondo temperature in
any case is a crossover scale, whose prefactor is a matter of convention.) A rough characteriza-
tion of the crossover from high to low temperatures is provided by the Curie-Weiss expression


χ(T ) ' 1/ 4(T +TKχ ) , but it should be recognized that this does not properly account for the
leading high- and low-T corrections indicated in Eq. (12).
Electron scattering: The rate at which conduction electrons scatter off the impurity has the
limiting forms
1
γ(T ) ∼ , T  TK , (13a)
ln(T /TK )
 
γ(T ) = γ(0) 1 − O(T /TK )2 , T /TK  1 . (13b)

At high temperatures, it is simply proportional to the effective coupling, g(T ). Its logarithmic
growth is cut off as screening sets in, and it approaches a constant as T → 0.
Fermi liquid behavior: Once the local spin is fully screened into a singlet, the “remaining”
conduction electrons scattering off it can no longer flip their spins; instead, they merely acquire
3.8 Jan von Delft

a phase shift. In this sense, the screened singlet acts as a static impurity, without any dynamics.
However, it is a strong scatterer: for an electron incident at the Fermi energy, the phase shift
acquired during scattering turns out to be δ(=0) = π/2, the maximum value possible for
scattering off static impurities.
Nozières has developed a Fermi-liquid description [13] of the strong-coupling regime. It in-
volves no dynamical impurity at all (since the latter is fully screened), only low-energy quasipar-
ticle excitations, i.e., dressed versions of the original, bare conduction electrons of the KM. The
quasiparticles experience mutual interactions that are weak and purely local, i.e., act only at the
site of the screened impurity. The parameters of the Hamiltonian are fully determined by the en-
ergy dependence of the scattering phase shift, δ(ε), and can be extracted from zero-temperature
properties of the KM (which in turn are known exactly from a Bethe Ansatz solution [5, 6]).
A perturbative treatment of these interactions readily yields the leading quadratic temperature
dependence, (T /TK )2 , indicated in Eqs. (12) for χ(T ) and (13a) for γ(T ). Fermi-liquid theory,
combined with results from the exact Bethe Ansatz solution of the Kondo model [5, 6], can also
be used to analytically derive the prefactors of the quadratic terms [14, 12], yielding

χ(T ) T2 γ(T ) π4 T 2
− 1 = −0.821 2 , −1=− 2
, T /TKχ  1 . (14)
χ(0) TKχ γ(0) 16 TKχ

These prefactors are universal constants, characteristic of the strong-coupling regime of the KM.
They serve as useful and stringent consistency checks for numerical treatments of the KM.

4 Numerical renormalization group

A description of the crossover from high to low temperatures, and of the strong-coupling regime,
requires nonperturbative approaches [3–10]. Here, we focus on the numerical renormalization
group (NRG), developed by Ken Wilson in 1975 as a specific implementation of his general RG
ideas [3]. Although almost 50 years old, NRG remains the gold standard for solving quantum
impurity models, due to its great flexibility, ability to resolve low-energy properties with high
accuracy, and the insights it yields about the RG evolution of a system from high to low energies.
Here, we discuss NRG for the example of the SIAM; since its low-energy behavior is governed
by the Kondo model, the SIAM, too, exhibits a Kondo effect, with an exponentially small Kondo
temperature. Hence, NRG results for the SIAM also serve to illustrate the physics for the KM
discussed in the previous section. The SIAM was first treated with NRG by Krishna-murthy,
Wilkens and Wilson in 1980 [15, 16]. For NRG reviews, see Refs. [4, 17].
Since the Kondo effect involves an exponentially small energy scale, Wilson’s goal was to de-
vise a numerical scheme capable of resolving arbitrarily low energy scales. He achieved this
goal by (i) discretizing the hybridization spectrum on a logarithmic grid that becomes arbitrarily
dense around the Fermi energy; (ii) mapping the resulting discrete model onto a semi-infinite
chain, called a Wilson chain, containing the impurity at the beginning and exponentially decay-
ing hopping matrix elements along the chain; and (iii) diagonalizing this chain iteratively, while
Impurity Models 3.9

Γ(ω ) = (∆/π )Θ(1 − |ω |)

(a)

−1 1 ω
(b) I−1 I−2 I+2 I+1

−1 −Λ−1 −Λ−2 Λ−2 Λ−1 1 ω

(c)

−1 ξ−1 ξ−2 ξ+2 ξ+1 1 ω

Fig. 1: (a) Box-shaped hybridization spectrum Γ (ω). (b) Partitioning of the support of Γ (ω)
into logarithmically spaced intervals, I±n . (c) Hybridization function after logarithmic dis-
cretization.

retaining only low-energy states. This procedure yields a set of approximate eigenstates of the
Hamiltonian spanning all energy scales of the model. We now discuss these steps one by one.
(i) Logarithmic discretization: The bath spectrum, defined on the interval  ∈ [−1, 1], is dis-
cretized on a logarithmic grid. Its grid points are chosen at ±Λ−n , with n = 1, 2, 3, . . . , N ,
where Λ > 1 is a discretization parameter, and N is large enough that Λ−N is smaller than the
smallest physical energy scale of interest (e.g. T ). (In practice, a typical choice is Λ = 2, and
N is chosen such that Λ−N is below machine precision, O(10−16 ).) The grid defines a set of
intervals I±n , defined as I+n = [Λ−n , Λ−n+1 ] and I−n = [−Λ−n+1 , Λ−n ] (see Fig. 1(b)). Next,
one represents each interval I±n by a single state, with energy ξ±n and hopping amplitude (to
the impurity) γ±n (see Fig. 1(c)), and replaces the bath Hamiltonian and hybridization terms by

X
star
Hbath → Hbath = ξ±n a†±ns a±ns , (15a)
±n,s
X
Hhyb → star
Hhyb = γ±n (a†±ns ds + h.c.) , (15b)
±n,s

where a±ns annihilates the representative bath mode ±n with spin s. This is known as the star
geometry, because the depiction of the impurity-bath couplings in Fig. 1 is reminiscent of rays
emanating from a star. If the discrete bath energies and couplings are chosen as [18],
Z R
2 I
dω Γ (ω)
γ±n = dω Γ (ω) , ξ±n = R ±n , (16)
I±n I±n
dω Γ (ω)/ω
P
then the resulting hybridization spectrum, Γ star () = ±n (γ± )2 δ(−ξ±n ), serves as a repre-
sentation of the original Γ () that becomes increasingly good the lower the energy [4]. (More
sophisticated choices are available to improve the approximation, see Refs. [19–21].)
3.10 Jan von Delft

star
(ii) Wilson chain: Next, the star geometry is mapped to a chain. To this end, note that in Hhyb ,
†
the impurity operator ds couples to discretized bath modes only in the combination td f0s =
P † †
±n γ±n a±ns . (td is chosen such that f0s obeys {f0s , f0s } = 1.) It is therefore advisable to
perform a unitary transformation from the 2N discrete modes {a±n,s } to a new set of L +1=2N
orthonormal modes containing f0s , say {f`s }, with ` = 0, 1, 2, . . . , L . This can be achieved
star star
by tridiagonalizing the single-particle Hamiltonian matrix for Hbath + Hhyb , say hstar , which has
dimensions (2N +1)×(2N +1), using Lanczos tridiagonalization. For example, for N = 2 we
obtain
   
ξ+2 0 γ+2 0 0 0 td 0 0 0
 0 ξ γ+1 0 0  t  t 0 0 
 +1  d 0 0 
   
hstar = γ+2 γ+1 0 γ−1 γ−2  → hchain
=  0 t 0 1 t 1 0 , (17)
   
 0 0 γ−1 ξ−1 0   0 0 t1 2 t2 
0 0 γ−2 0 ξ−2 0 0 0 t2 3

where the central row and column of hstar represent the mode ds , and the Lanczos scheme is
initialized with the vector (0, 0, 1, 0, 0)T . Such a tridiagonalization converts Hbath
star
into a form
representing a tight-binding chain of L +1 sites,
L
X −1 X L X
X
chain †

†
Hbath = t` f`s f`+1,s + h.c. + ` f`s f`s , (18a)
`=0 s `=0 s

with the impurity d level (at site ` = −1) coupled to site ` = 0 with amplitude td :
X

chain
Hhyb = td d†s f0s + h.c. . (18b)
s

The corresponding Wilson chain representation of the SIAM for a chain with largest site index
L is given by HL = Hloc + Hbath chain chain
+ Hhyb , with Hloc given by Eq. (1b).
An important consequence of logarithmic discretization is that the hopping matrix elements t`
and on-site energies ` decay exponentially along the chain: the integration intervals I±n in
Eqs. (16) have width Λ−n (Λ−1), thus γ±n ∼ Λ−n/2 and ξ±n ∼ Λ−n , which leads to t` ∼ Λ−`/2
and ` ∼ Λ−` . The characteristic energy scale of site ` is given by the hopping amplitude onto
that site, t`−1 ∼ Λ−(`−1)/2 . Thus, Wilson chains exhibit energy scale separation: different parts
of the chain represent different energy scales, with the characteristic energy of site ` decreasing
exponentially with increasing `.
(iii) Iterative diagonalization: The energy scale separation along a Wilson chain can be ex-
ploited to numerically diagonalize HL in iterative fashion, adding one site at a time, starting
from site ` = 0. In doing so, the resolution of the low-energy part of the eigenspectrum will
increase exponentially. However, the number of eigenstates will increase exponentially, too.
Therefore, a systematic truncation scheme is needed to limit the number of eigenstates that are
computed explicitly.
Wilson proposed an energy-based truncation scheme: Suppose that H ` has been diagonalized,
yielding a set of eigenstates satisfying H ` |αi` = Eα` |αi` . This set of eigenstates is called shell `.
Impurity Models 3.11

These are divided into two groups: the Nkeep lowest-lying states are kept, the remaining ones
discarded. (Typical choices for Nkeep range between 500 and 5000.) Then a new site, `+1, is
added, and H `+1 is diagonalized in the direct product space of the new site and the kept states of
shell `, whereas the contribution from its discarded states is neglected. Their neglect is justified
due to energy scale separation: the energies of the discarded, high-lying states from shell ` are
significantly larger than the characteristic energy of that shell, E`high  Λ−(`−1)/2 , which in turn
is larger than the coupling to the new site, t` = Λ−`/2 . The contribution of discarded states on
the eigenstates of the new shell L +1, estimated by a second-order perturbation argument, is of
order t2` /E`high. This is  (Λ−`/2 )2 /Λ−(`−1)/2 = Λ−(`+1)/2 , and hence much smaller than the
characteristic energy of the new shell, Λ−`/2 . Therefore, the effect of discarded states can be
safely neglected.
For a chain of total length `, Wilson’s iterative diagonalization scheme yields a set of Nkeep
eigenstates for each shell ` ≤ L . In combination, they resolve the spectrum of the Hamiltonian
on all energy scales from 1 to Λ−(L −1)/2 . Conceptually, the eigenenergies of shell ` represent
the low-energy part of the finite-size spectrum of the impurity plus bath put in a spherical box
of radius R` ∼ Λ(`−1)/2 , centered on the impurity. As ` increases, the box radius increases
exponentially, and the finite-size level spacing, 1/R` ∼ Λ−(`−1)/2 , decreases exponentially.
From an RG perspective, increasing R` amounts to decreasing the infrared cutoff, here given
by the finite-size level spacing, and thus probing the system at lower energy scales. A detailed
discussion of this perspective may be found in Refs. [3, 22].
To analyze how the structure of the spectrum changes with `, Wilson defined a set of rescaled
excitation energies for each shell,

Eα` = Λ(`−1)/2 Eα` − EGS
`
. (19)

These are measured with respect to the ground state (GS) energy of that shell, and scaled such
that the spacing of the lowest excitations is O(1). A plot of the rescaled energies Eα` versus shell
number ` on a linear scale, or equivalently, vs. the energy scale Λ−`/2 on a logarithmic scale,
with ` increasing in steps of 2, is called an NRG energy level flow diagram. It reveals how the
rescaled eigenspectra evolve (“flow”) with increasing system size. (` is increased in steps of 2
chain
because of an even-odd effect inherited from Hbath : its single-particle eigenspectra for L even
or odd are structurally different, causing the same to be true for the many-body eigenspectra
of HL.) An example of an NRG energy level flow diagram is shown in Fig. 2 below.
Once a Wilson chain has been iteratively diagonalized, the approximate eigenstates so obtained
can be used to compute physical quantities. These include thermodynamic quantities, such
as the impurity entropy S(T ) and spin susceptibility χ(T ), but also dynamical, frequency-
dependent quantities, such as the dynamic spin susceptibility χ(ω), or the impurity spectral
function, A(ω) and the retarded self-energy, Σ R (ω). For details, see Refs. [23, 24, 17].
3.12 Jan von Delft

Fig. 2: SIAM: NRG energy level flow diagrams (for ` = even), showing three fixed-point
regimes, associated with the free orbital (FO), local moment (LM) and strong-coupling (SC)
fixed points.

5 NRG results for the single-impurity Anderson model

In the following, we illustrate the RG flow of the SIAM with NRG results. Before we start, let
us specify our definition of the Kondo temperature for the SIAM. Following Haldane [25, 26]
and Refs. [14, 12], we use

1 p 2
π
TK = ' U ∆/2 e−πU/8∆+π∆/2U ex , x = εd + U/2 √ . (20)
4χ(0) 2U ∆

The first equality for TK corresponds to TKχ from our discussion of the spin susceptibility of the
Kondo model; the second follows from analytical results for χ(0) from the exact Bethe-Ansatz
solution of the Anderson model [5]. The parameter x measures the distance to the particle-hole
symmetric point, where εd = −U/2.
Or numerical results were computed using the following parameters (all energies are given in
units of D): U = 2 · 10−3 , εd = −U/2 = −10−3 , ∆ = 0.04 U, and h = 0 (unless stated
otherwise). We are thus in the wide-band limit, ∆, U  D = 1. The corresponding Kondo
temperature from Eq. (20) is TK = 1.3533 · 10−8 . The numerical NRG parameters were chosen
as Λ = 1.7, Nkeep = 5000.
Figure 2 shows an energy level flow diagram for the SIAM. We observe that there are three
energy ranges during which the flow of the finite-size spectrum is almost stationary (` indepen-
dent). Within each of these, Λ−`/2 lies within an energy regime governed by one of the RG
fixed points of the model. Here, the nomenclature “fixed point” really is appropriate: the spec-
trum literally remains fixed while changing ` (conceptually, increasing the system size R` ). The
energy ranges in between, where the levels shift and/or cross each other, are associated with
crossovers between fixed points.
Impurity Models 3.13

Fig. 3: SIAM: Top: Temperature dependence of the impurity entropy, S(T ). The plateaus at
ln(4) = 2 ln(2), ln(2) and ln(1) = 0 in the FO, LM and SC regimes, respectively, reflect the
number of states that are energetically available to the impurity in these regimes: four (|0i, |↑i,
|↓i, |↑↓i) in the FO regime, two (|↓i, |↑↓i) in the LM regime, and one (screened singlet) in the
SC regime. Bottom: Temperature dependence of the static spin susceptibility, χ(T ). In the LM
regime it shows Curie behavior, χ(T ) ∼ 1/T . In the SC regime it saturates at a value, χ(0),
which we use as definition for the Kondo temperature, setting TK = 1/(4χ(0)).

For the SIAM, the RG flow has three fixed point regimes (first analyzed with NRG in Refs. [15,
16]), clearly visible in Fig. 2. (I) The free orbital (FO) regime involves excitation energies & U.
It is governed by “free” (not hindered by U ) charge fluctuations, arising through real transitions
between all four states of the d level, |0i, |↑i, |↓i, and |↑↓i. (II) The local moment regime (LM)
involves excitation energies in the range between the Kondo temperature, TK and U. Here,
charge fluctuations can occur only virtually, giving rise to spin-flip transitions between |↑i and
| ↓i. Real transitions to |0i or | ↑↓i are frozen out, since the available excitation energies are
too small to overcome the associated cost in Coulomb energy, U. (III) The strong-coupling
(SC) regime involves excitation energies well below TK . It features a screened spin-singlet,
off which Fermi-liquid quasiparticles scatter without spin-flip scattering (cf. the discussion on
p. 7). Spin-flip scattering is frozen out, since that would require breaking up the spin singlet,
which has binding energy ∼ TK .
This interpretation of the fixed point spectra is beautifully confirmed by the temperature depen-
dence of two thermodynamic quantities, the impurity entropy, S(T ), and the static local spin
susceptibility, S(T ), shown in Fig. 3. As explained in the caption of these plots, both quanti-
ties show distinctly different behaviors in the three fixed point regimes, characteristic of a free
orbital, a local moment or a screened singlet, respectively.
3.14 Jan von Delft

Fig. 4: SIAM: Frequency dependence of the imaginary part of the dynamical spin susceptibility,
χ00 (ω), plotted on a log-log scale, for a range of different temperatures.

Next, we turn to dynamical quantities. First, we consider the dynamical spin susceptibility,
defined as
Z ∞
0 00
 
χ(ω) = χ (ω) + iχ (ω) = dt eiωt (−i) Θ(t) Sz (t), Sz (0) T . (21)
−∞

The static spin susceptibility discussed earlier, χ(T ), is equal to the zero-frequency limit of the
real part here, χ0 (ω = 0). The frequency dependence of the imaginary part, χ00 (ω), is shown
in Fig. 4. For T = 0 and ω  TK , it increases with decreasing ω, characteristic of an almost
free spin in the LM regime. It reaches a maximum at ω ' TK , where spin screening becomes
strong, and for ω → 0 decreases linearly with ω. This linear dependence is characteristic of
the Fermi liquid behavior of the system in the SC regime (energies  TK ), and indicative of a
well-screened impurity spin. For temperatures T  TK , the maximum in χ00 (ω) occurs at T,
and the peak value of χ00 decreases with increasing T. This illustrates how a finite temperature
cuts off the RG flow towards strong coupling.
Another dynamical quantity of great interest is the retarded correlator for the d level,
Z ∞ Z
As (ω 0 )
Gs (ω) = dt e (−i) Θ(t) {ds (t), ds (0)} T = dω 0
iωt
. (22)
−∞ ω − ω 0 + i0+

Specifically, we are interested in its spectral function, As (ω) = − π1 ImGs (ω), which charac-
terizes the local density of states associated with the d level, and in the corresponding retarded
self-energy, Σs (ω), defined via
1
Gs (ω) = , (23)
ω − εd − ∆(ω) − Σs (ω)
Impurity Models 3.15

Fig. 5: SIAM: Local spectral function, A(ω), for eight different temperatures, shown on a linear
scale (left) and a logarithmic scale (right).

Fig. 6: Local self-energy, Σ(ω), for eight different temperatures, shown on a linear scale (left)
and a logarithmic scale (right).

where ∆(ω) is the hybridization function from Eq. (2). Note that ∆(0) = −i∆, which is why
∆ sets the hybridization-induced width of the d level. (For the wide-band limit (∆, U  D)
considered here, ∆(ω) ' −i∆ actually holds to very good approximation throughout the regime
|ω| ≤ U/2.) Since we will mostly consider the case h = 0 (no magnetic field), we will
henceforth drop the spin index s.
Figure 5 shows the local spectral function, A(ω), for a number of different temperatures, rang-
ing from T  TK to T  TK . It features two broad peaks at ω = ±U/2, the so-called Hubbard
side-bands. These arising from real transitions between the d level and the band and are es-
sentially independent of T. Additionally, once T drops below TK , A(ω) develops a sharp peak
at ω = 0, known as the Kondo resonance. At zero temperature, its width is essentially given
3.16 Jan von Delft

Fig. 7: SIAM: Fermi-liquid scaling behavior of the local self-energy, Σ(ω, T ). For ω, T  TK ,
the rescaled data collapses onto a universal scaling curve in accordance with Eq. (24).

by TK and its height is A(0) = 1/(π∆). The latter relation follows from Eq. (23) by using
ImΣ(0) T =0 = 0, a well-established analytical result that follows from Nozières’ Fermi-liquid
treatment of the SC regime [13, 14, 12]. Figure 6, which shows the imaginary part of the self-
energy, −ImΣ(ω), confirms this fact: At T = 0, the self-energy indeed drops to zero in the
energy range |ω|  TK .
The occurrence of the Kondo resonance is perhaps the most striking manifestation of spin
screening. Intuitively, it arises because at very low energies, ω, T  TK , Fermi-liquid quasi-
particles scatter purely elastically off the screened singlet: they acquire a phase shift, but their
energy and spin does not change.
When ω and/or T increase from 0, inelastic scattering sets in and −ImΣ becomes nonzero. As
long as ω, T  TK , this can still be described using Fermi-liquid theory. By combining known
relations from Refs. [27, 12] one finds


π2∆ ω2+ π2T 2
−ImΣ(ω, T ) = , ω, T  TK . (24)
32TK2
This implies that a plot of −ImΣ(ω, T )/T 2 vs. ω/T should yield a scaling collapse. Indeed,
when the self-energy is computed using state-of-the-art NRG methodology [28,29], this scaling
behavior is beautifully recovered, as shown in Fig. 7.
The application of a local magnetic field h favors the state |↑i over |↓i. It thus tends to polarize
the local spin, thereby disrupting spin screening. As a result, the RG flow to strong coupling
is cut off at the energy scale h. This has dramatic consequences for the local spectral function:
The Kondo resonances for A↑ (ω) and A↓ (ω) shift apart and get suppressed with increasing
field. This behavior is shown in Fig. 8.
Impurity Models 3.17

Fig. 8: SIAM: Dependence of the local spectral function on magnetic field, computed at T = 0.

We conclude our discussion of the SIAM by putting it into the context of dynamical mean-field
theory (DMFT) [30]. There, one considers a correlated lattice model but assumes the self-
energy to be purely local. This allows the model to be mapped onto a quantum impurity model
with a hybridization function that has to be determined self-consistently. For the one-band
Hubbard model, the corresponding impurity model is the SIAM. DMFT self-consistency then
has drastic consequences: at sufficiently high temperature or interaction strength, the Kondo
resonance is completely suppressed [31, 32], and the system undergoes a Mott transition from a
metallic state to an insulating state.

6 Two-channel Kondo model

We end these notes by making a few remarks about a quantum impurity model that does not
show spin screening: the two-channel Kondo (2CK) model [33]. Its Hamiltonian is similar to
that of the usual (one-channel) Kondo model, but now there are two conduction bands, labeled
by a channel index j = 1, 2, that couple symmetrically to the local spin- 21 :
XX X X †
H2CK = εk n̂kjs + J Ŝd · ŝc , ŝc = ckjs 21 σss0 ck0 js0 . (25)
ks j=1,2 ks,k0 s0 j=1,2

This model also involves some screening of the local spin by conduction electrons, but full spin-
screening is not possible: if the local spin- 21 would form a spin singlet with a Kondo cloud from
one channel, that would break channel symmetry. If two Kondo clouds from both channels try
to screen the spin- 21 without breaking channel symmetry, they together constitute a spin-1 cloud
and overscreen the spin- 21 , yielding another spin- 21 object. Repeating this heuristic argument,
one concludes that a spin singlet cannot be formed as long as channel symmetry is respected.
Therefore, the quasiparticles do not behave as a Fermi liquid—instead, the model shows non-
Fermi-liquid behavior.
This has striking consequences for physical quantities. Here, we only mention two zero-
temperature properties that are decidedly non-Fermi-liquid in character. Figure 9 shows cor-
responding NRG results.
3.18 Jan von Delft

Fig. 9: Two-channel Kondo model. Top: temperature dependence
of the impurity entropy, S(T ).
We defined TK through the condition S(TK ) = S(∞)+S(0) /2 such that the crossover is
centered on TK . Bottom: Frequency dependence of the imaginary part of the dynamical spin
susceptibility, χ00 (ω), plotted on a log-log scale, for a range of different temperatures.

Impurity entropy: At T = 0, the impurity entropy S(0) does not equal ln(1) = 0, but 12 ln(2) =
√
ln( 2). The unusual fractional prefactor reflects the fact that the impurity retains some dynam-
ical character, since it is not fully screened.
Dynamical spin susceptibility: At T = 0, the imaginary part of the susceptibility approaches a
constant as ω → 0, χ00 (ω) ∼ const (in contrast to the Fermi liquid behavior, χ00 (ω) ∼ ω, of
the one-channel Kondo model). That, in turn, implies that the real part diverges logarithmically,
χ0 (ω) ∼ ln(TK /ω). This, too, indicates that the impurity is not fully screened but retains some
dynamical character, even at T = 0.

Acknowledgement

I warmly thank Andreas Gleis and Jeongmin Shim for doing the NRG computations reported in
these notes and creating the corresponding figures, and Andreas Gleis for establishing Eq. (24).
Impurity Models 3.19

A Schrieffer-Wolff transformation
The KM can be obtained from the SIAM using a Schrieffer-Wolff transformation [11] and a pro-
jection onto the subspace were nd = 1. Here, we briefly outline the key steps of this derivation.
The starting point is the SIAM Hamiltonian, written in the form HSIAM = H0 + H1 , with
H0 = Hbath + Hloc being O(v 0 ) and H1 = Hhyb being O(v 1 ), where v = vk is the hybridization
parameter. The goal is to find a unitary transformation such that the new Hamiltonian, H e =
eA HSIAM e−A , contains no terms of order O(v 1 ). Here, A is anti-unitary, A† = −A, with an
e in powers of A
hybridization expansion of the form A = O(v 1 ) + O(v 2 ) + . . . . Expanding H
using the Baker-Campbell-Hausdorff formula, one obtains
 . 
e
H = (H0 +H1 ) + [A, H0 +H1 ] + 2 A, [A, H0 +H1 ] + O(v 3 )
1
(26)
(The term slashed out is O(v 3 ).) To ensure that the right side contains no terms O(v 1 ), we must
choose A such that H1 = −[A, H0 ], implying
e = H0 + 1 [A, H1 ] + O(v 3 ) .
H (27)
2

One may readily verify that the requirement H1 = −[A, H0 ] is satisfied by

X  1 † U † †

A= v cks ds + d−s d−s cks ds − h.c. (28)
ε k −εd (ε d − εk )(εd + U − εk )
ks
Inserting this into Eq. (27) and dropping terms that are zero in the subspace where nd = 1, one
obtains
X X
e n =1 =
H| ε n̂
k ks + vekk0 Ŝd · c†ks 21 σss0 ck0 s0 + . . . (29)
d
ks kk0
where the coupling matrix elements are given by
− 12 v 2 U
ve =
kk0 + (k ↔ k 0 ) . (30)
(εd − εk )(εd + U − εk )
P
In Eq. (29), the ellipsis . . . stands for terms of the form kk0 c†ks ck0 s , describing potential scat-
tering, i.e., scattering without any spin dependence. These are usually ignored when discussing
the Kondo effect, because in contrast to the spin-exchange interaction of the second term in
Eq. (29), they do not give rise to a logarithmic temperature dependence when treated pertur-
batively. In renormalization group (RG) terminology, spin-exchange is a relevant perturbation,
whereas potential scattering is not.
Now consider low-energy excitations close to the Fermi energy, then vekk0 simplifies to
v2U
vekk0 ' − = J, ∀|εk |, |εk0 |  |εd |, |εd +U | . (31)
εd (εd + U )
e of Eq. (29) thus reduces to the Kondo Hamiltonian HKM of Eq. (7), with an
In this limit, H
exchange coupling constant J given by Eq. (31). Note that this equals the expression for J
given in Eq. (6) above, found by considering 2nd order spin-flip processes.
To summarize the above discussion: the low-energy behavior of the SIAM, at energy scales for
which charge fluctuations involving |0i and |↑↓i are ”frozen out” (can occur only virtually), is
described by the KM.
3.20 Jan von Delft

References
[1] W.J. de Haas, J. de Boer, and G.J. van den Berg, Physica 1, 1115 (1934)

[2] J. Kondo, Prog. Theor. Phys. 32, 37 (1964)

[3] K.G. Wilson, Rev. Mod. Phys. 47, 773 (1975)

[4] R. Bulla, T.A. Costi, and T. Pruschke, Rev. Mod. Phys. 80, 395 (2008)

[5] A. Tsvelick and P. Wiegmann, Adv. Phys. 32, 453 (1983)

[6] N. Andrei, K. Furuya, and J.H. Lowenstein, Rev. Mod. Phys. 55, 331 (1983)

[7] I. Affleck and A. Ludwig, Nucl. Phys. B 352, 849 (1991)

[8] V.J. Emery and S. Kivelson, Phys. Rev. B 46, 10812 (1992)

[9] G. Zaránd and J. von Delft, Phys. Rev. B 61, 6918 (2000)

[10] E. Gull, A.J. Millis, A.I. Lichtenstein, A.N. Rubtsov, M. Troyer, and P. Werner,
Rev. Mod. Phys. 83, 349 (2011)

[11] J.R. Schrieffer and P.A. Wolff, Phys. Rev. 149, 491 (1966)

[12] M. Filippone, P. Moca, A. Weichselbaum, J. von Delft, and C. Mora,

Phys. Rev. B 98, 075404 (2018)

[13] P. Nozières, J. Low Temp. Phys. 17, 31 (1974)

[14] C. Mora, C.P. Moca, J. von Delft, and G. Zaránd, Phys. Rev. B 92, 075120 (2015)

[15] H.R. Krishna-murthy, J.W. Wilkins, and K.G. Wilson, Phys. Rev. B 21, 1003 (1980)

[16] H.R. Krishna-murthy, J.W. Wilkins, and K.G. Wilson, Phys. Rev. B 21, 1044 (1980)

[17] A. Weichselbaum, Phys. Rev. B 86, 245124 (2012)

[18] J. Vivaldo L. Campo and L.N. Oliveira, Phys. Rev. B 72, 104432 (2005)

[19] R. Žitko, Comp. Phys. Commun. 180, 1271 (2009)

[20] R. Žitko and T. Pruschke, Phys. Rev. B 79, 085106 (2009)

[21] B. Bruognolo, N.O. Linden, F. Schwarz, S.S.B. Lee, K. Stadler, A. Weichselbaum,

M. Vojta, F.B. Anders, and J. von Delft, Phys. Rev. B 95, 121115 (2017)

[22] J. von Delft, G. Zaránd, and M. Fabrizio, Phys. Rev. Lett. 81, 196 (1998)

[23] R. Peters, T. Pruschke, and F.B. Anders, Phys. Rev. B 74, 245114 (2006)
Impurity Models 3.21

[24] A. Weichselbaum and J. von Delft, Phys. Rev. Lett. 99, 076402 (2007)

[25] F.D.M. Haldane, Phys. Rev. Lett. 40, 416 (1978)

[26] F.D.M. Haldane, J. Phys. C: Solid State Phys. 11, 5015 (1978)

[27] A.C. Hewson, J. Phys.: Condens. Matter 13, 10011 (2001)

[28] F.B. Kugler, Phys. Rev. B 105, 245132 (2022)

[29] A. Gleis (unpublished) (2022)

[30] A. Georges, G. Kotliar, W. Krauth, and M.J. Rozenberg, Rev. Mod. Phys. 68, 13 (1996)

[31] R. Bulla, Phys. Rev. Lett. 83, 136 (1999)

[32] R. Bulla, T.A. Costi, and D. Vollhardt, Phys. Rev. B 64, 045103 (2001)

[33] P. Nozières and P. Blandin, J. Phys. 41, 193 (1980)