Vous êtes sur la page 1sur 9

Indian Institute Of Technology

Kharagpur

Exact Diagonalization of the Hubbard


Model

Advanced Computational Physics


Arnab Barman Ray
13PH20011

April 6, 2017
1 Introduction
The Hubbard model, which was first introduced by John Hubbard in 1963
has still remained an area of active research to this date. Variations of this
simple model has found applications in explaining features of high temper-
ature superconductivity, anti-ferromagnetism and other strongly correlated
phenomena. The original Hamiltonian is simplistic in nature and consists
of two terms: the on-site repulsion represented by U which signifies the re-
pulsion between two electrons of opposite spins on the same lattice site, and
another hopping term, represented by t which is a measure of overlap of or-
bitals between neighbouring sites. The hamiltonian is presented in equation
(1): X X
H = t (ci cj ) + U ni ni (1)
<i,j> i

n = c c counts the number of electrons in the particular state in any


wavefunction . Each site i in the lattice can hold 2 electrons at most with
differing spin directions. This Hamiltonian as can be seen allows no admix of
up-pin and down-spin electrons. It can P also be shown that it commutes with
total particle number operator, N = i, ni, which counts the totalP number
P in the system, as well as the number operators N = i ni, and
of particles
N = i ni, , which counts the number of up-spin and sown-spin electrons
respectively.
This incredible property of the Hamiltonian guarantees us that any eigen-
state would have a particular number of up-spin as well as a particular num-
ber of down-spin electrons. Hence, physicists are given the freedom to operate
in any subspace of the total Hilbert space which is incredibly huge ( 2N for
N states).Essentially we have the freedom to choose a smaller set of basis
having a particular number of electrons.
The half-filled case is where the number of electrons in the system is
the same as the number of sites, irrespective of spin.It has been of particular
interest since it represents a natural situation where each atom(site) provides
one electron to the system. The Hamiltonian too supposedly has some nice
mathematical properties for this case.
In this report we present an analysis of the ground state wavefunction for
the half-filled case with the periodic boundary condition for N = 4, 6, 8 and
10 electrons(sites) using the Lanczos method to diagonalize the Hamiltonian.
We also investigate it to see if the obtained state is anti-ferromagnetic.

1
Sz State I I I = 2N I + I
2 |00i |11i 3 0 3
0 |01i |01i 1 1 5
0 |01i |10i 3 1 7
0 |10i |01i 1 2 9
0 |10i |10i 2 2 10
-2 |11i |00i 3 0 12

Table 1: Basis states for N = 2


Sites Dimension
4 70
9 924
8 12870
10 184756
Table 2: Dimensions involved.

2 The Hamiltonian Matrix


As an example, the method of organising the basis states for the case of
N = 2 is presented in the Table 1.
Such an ordering was first introduced in [1].The property that the spin up
and spin down number operator commutes with the Hamiltonian allows us
to segregate the basis into sectors based on their total spin component Sz as
is shown in the table(here, there are three sectors).The states are separated
for the spins are represented in binary notation where 0 means unoccupied
and 1 means occupied for a particular site.These binary representations are
then converted into numbers.
The Hamiltonian is block-diagonal in each sector.There is an unique iden-
tification number I associated with each basis state. Within a sector this
number is sorted, however one must be careful that unlike in this example,
in a larger system smaller values of I might occur in a later sector.For N
sites there will be in total N + 1 spin sectors with each i-th sector having
N!
(N i)!i!
states, where i is the number of electrons with spin up/down.The
number of bases for each case is illustrated in Table 2.As can be seen, this
increase is exponential.Even with modern supercomputers, atmost 20 site
calculations can be achieved. The interaction part of the Hamiltonian was

2
easy to construct and an element can be initialized by just taking the AND
operator of the two binary representations of I(between the two states) and
counting the number of 1s in the resulting binary number.
The hopping part is a bit more difficult, especially with the periodic
boundary condition.A modification of a python library constructed by Kim
G. L. Pedersen for an extended system was used.

3 Lanczos method
The Lanczos method of iteration provides an easy way of calculating the
eigenvalues and eigenvectors of large matrices without the need for high
computational power.However, it is also an unstable procedure because the
vectors that are generated by the process quickly loose orthogonality and
are to be re-orthogonalized after a certain number of iterations which is
machine-dependent.These operations cause more time to be consumed than
the matrix-vector multiplication that is the main step in the algorithm.
The basic idea in the Lanczos method is to form a tri-diagonal ma-
trix in a new basis of something called a Krylov space, where the vectors
v0 , Hv0 , H 2 v0 , ... reside, where H is the Hamiltonian and v0 is an arbitar-
ily chosen random vector.The basic algorithm is illustrated in a pseudocode
below[4]:
f u n c t i o n [ alpha , beta ] = l e c 3 2 l a n c z o s (A, q1 , b t o l , kmax ) ;
%
% Run a b a s i c Lanczos i t e r a t i o n u n t i l e i t h e r b e t a k < b t o l
% o r k == kmax .
k = 0;
qk = 0 ;
r = q1 ;
b = 1;
w h i l e ( b > b t o l ) & ( k < kmax )
k = k+1;
qkm1 = qk ;
qk = r /b ;
Aqk = Aqk ;
a lpha ( k ) = qkTAqk ;
r = Aqkqk alpha ( k)qkm1b ;
b = norm ( r ) ;

3
b e ta ( k ) = b ;
end
The tridiagonal matrix T generated from the coefficients for k iterations
has the form:

1 1 0 0...
1 2 2 0 . . .

0 0...
T =


..
.
0 0 . . . k1 k k
The eigenvalues of this matrix approaches that of the original matrix for
large enough values of k. The eigenvectors are then constructed using the
basis vectors and T .We have used 200 iterations in our program to calculate
the eigenvalue and the corresponding eigenvectors.

4 Ground states
Since the number of dimensions increases rapidly, to show the ground state,
we have chosen to represent it as the combined probability of the wavefunc-
tion to be in one particular Sz sector.Hence, a point includes the sum of
the probabilities of all the states in that particular sector.The three cases
are shown in Fig. 1 along with their respective energies. The value of the
hopping potential is 1 while the on-site repulsion is 5.
As we can see in all the three cases, the wavefunction is severly restricted
to one spin sector. This is due to the fact that the N and N operators and
hence with Sz = N N commutes with H, meaning that any eigenvector of
the Hamiltonian will also be an eigenvector of the Sz and will hence contained
in a specific spin state.
Unlike for the first three cases, the wavefunction for N = 10 is not as
expected due to smaller number of iterations made on account of low mem-
ory.Hence, we have not included that case in this report.
The impartiality of the Hamiltonian towards up or down spin is also
illustrated in the fact that the wavefunction is perfectly symmetric.

4
Figure 1: Eigenvalue= -1.657

Figure 2: Eigenvalue= -2.750

Figure 3: Eigenvalue= -3.844

5
5 Anti-ferromagnetism
One of the early successes of the Hubbard model was in explaining the
insulator-conductor transition in certain materials.It has also been shown
to host a variety of non-trivial phenomena such as ferromagnetism, anti-
ferromagnetism etc. In our case, for this particular choice of parameters, we
expect that the ground states will be anti-ferromagnetic in the sense that ba-
sis states which have only one spin on each site and where neighbouring spins
have different signs, will have maximum contribution to the eigenfunction.
To investigate the ground state, we essentially blow-up the Sz = 0 sector
of the wavefunction in the following figures. As we can see there are two
prominent peaks in all the cases.
For N = 4, the two peaks are close to the anti-ferromagnetic labels of
165 and 60:

Figure 4: N = 4

Similarly, for N = 6, we get to see peaks in the wavefunction near the


antiferromagnetic labels of 2709 and 1386.
In the case of N = 8, the two peaks correspond to the states, h|
h01010101| h10101010| and h| h10101010| h10101010| . Once
can easily verify that these states have labels of 21930 and 43605, which are
in good agreement with the figure.
Hence, we can conclude in good faith that these ground states are largely
anti-ferromagnetic in nature.

6
Figure 5: N = 6

Figure 6: N = 8

6 Conclusion
The ground state eigenfunctions and eigenvalues of the half-filled Hubbard
Model were calculated for 4, 6, 8 and 10 sites using the Lanczos iteration
algorithm for parameter values of t = 1 and U = 5.
The anti-ferromagnetism property of these states was verified for each
case except N = 10 where due to memory issues, the number of iterations
had to be restricted leading to a non-characteristic wavefunction which isnt
centered in just one spin sector.

6.1 References
The following is a list of materials I had consulted to carry out the project:

1. Introduction to Hubbard Model and Exact Diagonalization - SA Jafari

2. The Lanczos Method - Erich Koch

7
3. Exact diagonalization of the Hubbard model on graphics processing
units- T.Siro, A.Harju

4. Cornell CS9210 Lecture notes

Vous aimerez peut-être aussi