2.9. Calculating spectral density with the kernel polynomial method

We have already seen in the “Closed systems” tutorial that we can use Kwant simply to build Hamiltonians, which we can then directly diagonalize using routines from Scipy.

This already allows us to treat systems with a few thousand sites without too many problems. For larger systems one is often not so interested in the exact eigenenergies and eigenstates, but more in the density of states.

The kernel polynomial method (KPM), is an algorithm to obtain a polynomial expansion of the density of states. It can also be used to calculate the spectral density of arbitrary operators. Kwant has an implementation of the KPM method kwant.kpm, that is based on the algorithms presented in Ref. 1.

See also

The complete source code of this example can be found in kernel_polynomial_method.py

Introduction

Our aim is to use the kernel polynomial method to obtain the spectral density \(ρ_A(E)\), as a function of the energy \(E\), of some Hilbert space operator \(A\). We define

\[ρ_A(E) = ρ(E) A(E),\]

where \(A(E)\) is the expectation value of \(A\) for all the eigenstates of the Hamiltonian with energy \(E\), and the density of states is

\[ρ(E) = \sum_{k=0}^{D-1} δ(E-E_k) = \mathrm{Tr}\left(\delta(E-H)\right),\]

where \(H\) is the Hamiltonian of the system, \(D\) the Hilbert space dimension, and \(E_k\) the eigenvalues.

In the special case when \(A\) is the identity, then \(ρ_A(E)\) is simply \(ρ(E)\), the density of states.

Calculating the density of states

Roughly speaking, KPM approximates the density of states, or any other spectral density, by expanding the action of the Hamiltonian and operator of interest on a small set of random vectors (or local vectors for local density of states), as a sum of Chebyshev polynomials up to some order, and then averaging. The accuracy of the method can be tuned by modifying the order of the Chebyshev expansion and the number of vectors. See notes on accuracy below for details.

Performance and accuracy

The KPM method is especially well suited for large systems, and in the case when one is not interested in individual eigenvalues, but rather in obtaining an approximate spectral density.

The accuracy in the energy resolution is dominated by the number of moments. The lowest accuracy is at the center of the spectrum, while slightly higher accuracy is obtained at the edges of the spectrum. If we use the KPM method (with the Jackson kernel, see Ref. 1) to describe a delta peak at the center of the spectrum, we will obtain a function similar to a Gaussian of width \(σ=πa/N\), where \(N\) is the number of moments, and \(a\) is the width of the spectrum.

On the other hand, the random vectors will explore the range of the spectrum, and as the system gets bigger, the number of random vectors that are necessary to sample the whole spectrum reduces. Thus, a small number of random vectors is in general enough, and increasing this number will not result in a visible improvement of the approximation.

The global spectral density \(ρ_A(E)\) is approximated by the stochastic trace, the average expectation value of random vectors \(r\)

\[ρ_A(E) = \mathrm{Tr}\left(A\delta(E-H)\right) \sim \frac{1}{R} \sum_r \langle r \lvert A \delta(E-H) \rvert r \rangle,\]

while the local spectral density for a site \(i\) is

\[ρ^i_A(E) = \langle i \lvert A \delta(E-H) \rvert i \rangle,\]

which is an exact expression.

Global spectral densities using random vectors

In the following example, we will use the KPM implementation in Kwant to obtain the (global) density of states of a graphene disk.

We start by importing kwant and defining our system.

# necessary imports
import kwant
import numpy as np


# define the system
def make_syst(r=30, t=-1, a=1):
    syst = kwant.Builder()
    lat = kwant.lattice.honeycomb(a, norbs=1)

    def circle(pos):
        x, y = pos
        return x ** 2 + y ** 2 < r ** 2

    syst[lat.shape(circle, (0, 0))] = 0.
    syst[lat.neighbors()] = t
    syst.eradicate_dangling()

    return syst

After making a system we can then create a SpectralDensity object that represents the density of states for this system.

    fsyst = make_syst().finalized()

    spectrum = kwant.kpm.SpectralDensity(fsyst, rng=0)

The SpectralDensity can then be called like a function to obtain a sequence of energies in the spectrum of the Hamiltonian, and the corresponding density of states at these energies.

    energies, densities = spectrum()

When called with no arguments, an optimal set of energies is chosen (these are not evenly distributed over the spectrum, see Ref. 1 for details), however it is also possible to provide an explicit sequence of energies at which to evaluate the density of states.

    energy_subset = np.linspace(0, 2)
    density_subset = spectrum(energy_subset)
../_images/kpm_dos.png

In addition to being called like functions, SpectralDensity objects also have a method integrate which can be used to integrate the density of states against some distribution function over the whole spectrum. If no distribution function is specified, then the uniform distribution is used:

    print('identity resolution:', spectrum.integrate())
identity resolution: (6492.999999999999+1.5408785127112895e-30j)

We see that the integral of the density of states is normalized to the total number of available states in the system. If we wish to calculate, say, the number of states populated in equilibrium, then we should integrate with respect to a Fermi-Dirac distribution:

    # Fermi energy 0.1 and temperature 0.2
    fermi = lambda E: 1 / (np.exp((E - 0.1) / 0.2) + 1)

    print('number of filled states:', spectrum.integrate(fermi))
number of filled states: (3306.7989795427643-6.896600008816472e-15j)
Stability and performance: spectral bounds

The KPM method internally rescales the spectrum of the Hamiltonian to the interval (-1, 1) (see Ref. 1 for details), which requires calculating the boundaries of the spectrum (using scipy.sparse.linalg.eigsh). This can be very costly for large systems, so it is possible to pass this explicitly as via the bounds parameter when instantiating the SpectralDensity (see the class documentation for details).

Additionally, SpectralDensity accepts a parameter epsilon, which ensures that the rescaled Hamiltonian (used internally), always has a spectrum strictly contained in the interval (-1, 1). If bounds are not provided then the tolerance on the bounds calculated with scipy.sparse.linalg.eigsh is set to epsilon/2.

Local spectral densities using local vectors

The local density of states can be obtained without using random vectors, and using local vectors instead. This approach is best when we want to estimate the local density on a small number of sites of the system. The accuracy of this approach depends only on the number of moments, but the computational cost increases linearly with the number of sites sampled.

To output local densities for each local vector, and not the average, we set the parameter mean=False, and the local vectors will be created with the LocalVectors generator (see advanced_topics for details).

The spectral density can be restricted to the expectation value inside a region of the system by passing a where function or list of sites to the RandomVectors or LocalVectors generators.

In the following example, we compute the local density of states at the center of the graphene disk, and we add a staggering potential between the two sublattices.

# define the system
def make_syst_staggered(r=30, t=-1, a=1, m=0.1):
    syst = kwant.Builder()
    lat = kwant.lattice.honeycomb(a, norbs=1)

    def circle(pos):
        x, y = pos
        return x ** 2 + y ** 2 < r ** 2

    syst[lat.a.shape(circle, (0, 0))] = m
    syst[lat.b.shape(circle, (0, 0))] = -m
    syst[lat.neighbors()] = t
    syst.eradicate_dangling()

    return syst

Next, we choose one site of each sublattice “A” and “B”,

    # find 'A' and 'B' sites in the unit cell at the center of the disk
    center_tag = np.array([0, 0])
    where = lambda s: s.tag == center_tag
    # make local vectors
    vector_factory = kwant.kpm.LocalVectors(fsyst, where)

and plot their respective local density of states.

    # 'num_vectors' can be unspecified when using 'LocalVectors'
    local_dos = kwant.kpm.SpectralDensity(fsyst, num_vectors=None,
                                          vector_factory=vector_factory,
                                          mean=False,
                                          rng=0)
    energies, densities = local_dos()
    plot_dos([
        ('A sublattice', (energies, densities[:, 0])),
        ('B sublattice', (energies, densities[:, 1])),
    ])
../_images/kpm_ldos_sites.png

Note that there is no noise comming from the random vectors.

Increasing the accuracy of the approximation

SpectralDensity has two methods for increasing the accuracy of the method, each of which offers different levels of control over what exactly is changed.

The simplest way to obtain a more accurate solution is to use the add_moments method:

    spectrum.add_moments(energy_resolution=0.03)

This will update the number of calculated moments and also the default number of sampling points such that the maximum distance between successive energy points is energy_resolution (see notes on accuracy).

Alternatively, you can directly increase the number of moments with add_moments, or the number of random vectors with add_vectors. The added vectors will be generated with the vector_factory.

    spectrum.add_moments(100)
    spectrum.add_vectors(5)
../_images/kpm_dos_r.png

Calculating the spectral density of an operator

Above, we saw how to calculate the density of states by creating a SpectralDensity and passing it a finalized Kwant system. When instantiating a SpectralDensity we may optionally supply an operator in addition to the system. In this case it is the spectral density of the given operator that is calculated.

SpectralDensity accepts the operators in a few formats:

  • explicit matrices (numpy array or scipy sparse matrices will work)

  • operators from kwant.operator

If an explicit matrix is provided then it must have the same shape as the system Hamiltonian.

    # identity matrix
    matrix_op = scipy.sparse.eye(len(fsyst.sites))
    matrix_spectrum = kwant.kpm.SpectralDensity(fsyst, operator=matrix_op, rng=0)

Or, to do the same calculation using kwant.operator.Density:

    # 'sum=True' means we sum over all the sites
    kwant_op = kwant.operator.Density(fsyst, sum=True)
    operator_spectrum = kwant.kpm.SpectralDensity(fsyst, operator=kwant_op, rng=0)

Spectral density with random vectors

Using operators from kwant.operator allows us to calculate quantities such as the local density of states by telling the operator not to sum over all the sites of the system:

    # 'sum=False' is the default, but we include it explicitly here for clarity.
    kwant_op = kwant.operator.Density(fsyst, sum=False)
    local_dos = kwant.kpm.SpectralDensity(fsyst, operator=kwant_op, rng=0)

SpectralDensity will properly handle this vector output, and will average the local density obtained with random vectors.

The accuracy of this approximation depends on the number of random vectors used. This allows us to plot an approximate local density of states at different points in the spectrum:

    zero_energy_ldos = local_dos(energy=0)
    finite_energy_ldos = local_dos(energy=1)
    plot_ldos(fsyst, [
        ('energy = 0', zero_energy_ldos),
        ('energy = 1', finite_energy_ldos)
    ])
../_images/kpm_ldos.png

Calculating Kubo conductivity

The Kubo conductivity can be calculated for a closed system with two KPM expansions. In Conductivity we implemented the Kubo-Bastin formula of the conductivity and any temperature (see Ref. 2). With the help of Conductivity, we can calculate any element of the conductivity tensor \(σ_{αβ}\), that relates the applied electric field to the expected current.

\[j_α = σ_{α, β} E_β\]

In the following example, we will calculate the longitudinal conductivity \(σ_{xx}\) and the Hall conductivity \(σ_{xy}\), for the Haldane model. This model is the first and one of the most simple ones for a topological insulator.

The Haldane model consist of a honeycomb lattice, similar to graphene, with nearest neigbours hoppings. To turn it into a topological insulator we add imaginary second nearest neigbours hoppings, where the sign changes for each sublattice.

# define a Haldane system
def make_syst_topo(r=30, a=1, t=1, t2=0.5):
    syst = kwant.Builder()
    lat = kwant.lattice.honeycomb(a, norbs=1, name=['a', 'b'])

    def circle(pos):
        x, y = pos
        return x ** 2 + y ** 2 < r ** 2

    syst[lat.shape(circle, (0, 0))] = 0.
    syst[lat.neighbors()] = t
    # add second neighbours hoppings
    syst[lat.a.neighbors()] = 1j * t2
    syst[lat.b.neighbors()] = -1j * t2
    syst.eradicate_dangling()

    return lat, syst.finalized()

To calculate the bulk conductivity, we will select sites in the unit cell in the middle of the sample, and create a vector factory that outputs local vectors

    # construct the Haldane model
    lat, fsyst = make_syst_topo()
    # find 'A' and 'B' sites in the unit cell at the center of the disk
    where = lambda s: np.linalg.norm(s.pos) < 1

    # component 'xx'
    s_factory = kwant.kpm.LocalVectors(fsyst, where)
    cond_xx = kwant.kpm.conductivity(fsyst, alpha='x', beta='x', mean=True,
                                     num_vectors=None, vector_factory=s_factory,
                                     rng=0)
    # component 'xy'
    s_factory = kwant.kpm.LocalVectors(fsyst, where)
    cond_xy = kwant.kpm.conductivity(fsyst, alpha='x', beta='y', mean=True,
                                     num_vectors=None, vector_factory=s_factory,
                                     rng=0)

    energies = cond_xx.energies
    cond_array_xx = np.array([cond_xx(e, temperature=0.01) for e in energies])
    cond_array_xy = np.array([cond_xy(e, temperature=0.01) for e in energies])

    # area of the unit cell per site
    area_per_site = np.abs(np.cross(*lat.prim_vecs)) / len(lat.sublattices)
    cond_array_xx /= area_per_site
    cond_array_xy /= area_per_site

Note that the Kubo conductivity must be normalized with the area covered by the vectors. In this case, each local vector represents a site, and covers an area of half a unit cell, while the sum covers one unit cell. It is possible to use random vectors to get an average expectation value of the conductivity over large parts of the system. In this case, the area that normalizes the result, is the area covered by the random vectors.

../_images/kpm_cond.png

Advanced topics

Custom distributions of vectors

By default SpectralDensity will use random vectors whose components are unit complex numbers with phases drawn from a uniform distribution. The generator is accesible through RandomVectors.

For large systems, one will generally resort to random vectors to sample the Hilbert space of the system. There are several reasons why you may wish to make a different choice of distribution for your random vectors, for example to enforce certain symmetries or to only use real-valued vectors.

To change how the random vectors are generated, you need only specify a function that takes the dimension of the Hilbert space as a single parameter, and which returns a vector in that Hilbert space:

    # construct a generator of vectors with n random elements -1 or +1.
    n = fsyst.hamiltonian_submatrix(sparse=True).shape[0]
    def binary_vectors():
        while True:
           yield np.rint(np.random.random_sample(n)) * 2 - 1

    custom_factory = kwant.kpm.SpectralDensity(fsyst,
                                               vector_factory=binary_vectors(),
                                               rng=0)

Aditionally, a LocalVectors generator is also available, that returns local vectors that correspond to the sites passed. Note that the vectors generated match the sites in order, and will be exhausted after all vectors are drawn from the factory.

Both RandomVectors and LocalVectors take the argument where, that restricts the non zero values of the vectors generated to the sites in where.

Reproducible calculations

Because KPM internally uses random vectors, running the same calculation twice will not give bit-for-bit the same result. However, similarly to the funcions in rmt, the random number generator can be directly manipulated by passing a value to the rng parameter of SpectralDensity. rng can itself be a random number generator, or it may simply be a seed to pass to the numpy random number generator (that is used internally by default).

Defining operators as sesquilinear maps

Above, we showed how SpectralDensity can calculate the spectral density of operators, and how we can define operators by using kwant.operator. If you need even more flexibility, SpectralDensity will also accept a function as its operator parameter. This function must itself take two parameters, (bra, ket) and must return either a scalar or a one-dimensional array. In order to be meaningful the function must be a sesquilinear map, i.e. antilinear in its first argument, and linear in its second argument. Below, we compare two methods for computing the local density of states, one using kwant.operator.Density, and the other using a custom function.

    rho = kwant.operator.Density(fsyst, sum=True)

    # sesquilinear map that does the same thing as `rho`
    def rho_alt(bra, ket):
        return np.vdot(bra, ket)

    rho_spectrum = kwant.kpm.SpectralDensity(fsyst, operator=rho, rng=0)
    rho_alt_spectrum = kwant.kpm.SpectralDensity(fsyst, operator=rho_alt, rng=0)

References

1(1,2,3,4)

Rev. Mod. Phys., Vol. 78, No. 1 (2006).

2

Phys. Rev. Lett. 114, 116602 (2015).