Skip to content

Correlators and distributions

On this page we measure the ground state beyond a two-point correlator: four-point correlators, the two-particle density matrix, and the full distribution of an operator. Two-point correlators and the equal-time structure factor are on the Ground state page, and the single-particle correlation matrix, whose eigenvalues are the natural orbitals, is under Entanglement.

Four-point correlators

For a fermionic chain fc, the four-point tensor \(\langle c_i^\dagger c_jc_k^\dagger c_l\rangle\) for all \(i,j,k,l\) is

t = fc.get_gs().get_four_correlation_tensor()   # shape (n, n, n, n)

It holds every density-density correlator \(\langle n_in_k\rangle\) (at \(j=i\), \(l=k\)) and every pair correlator, so it is the input of a two-particle reduced density matrix or of a cumulant. The keyword ctmode chooses the method ("batched", "sweep", "fold", "full", "explicit"). The default chooses the fastest one for the backend, and "explicit" works on every backend. A single four-point correlator is also fc.vev(fc.Cdag[i]*fc.C[j]*fc.Cdag[k]*fc.C[l]), which is cheaper when only a few are needed.

The two-particle density matrix \(\langle c_i^\dagger c_j^\dagger c_lc_k\rangle\), with the pairs \(i<j\) and \(k<l\) as row and column, is

from dmrgpy.entropytk.correlationentropy import get_highorder_correlation_matrix
d2 = get_highorder_correlation_matrix(fc)

Operator distributions

For a Hermitian operator \(X\), the expectation value \(\langle X\rangle\) is only the mean. The full distribution of the measured values of \(X\) in the ground state is

\[P(x)=\langle\mathrm{GS}|\,\delta(X-x)\,|\mathrm{GS}\rangle,\qquad \int P(x)\,dx=1.\]

For a conserved charge this is the full counting statistics. A sharp \(P(x)\) means that the ground state has a well-defined value of \(X\), and a broad one means large fluctuations. It is computed with the kernel polynomial method, as the dynamical correlators:

X = sum(sc.Sz[i] for i in range(n))      # total magnetization
x, p = sc.get_distribution(X=X, delta=0.1)
mus, shift, scale = sc.get_distribution_moments(X=X, delta=0.1)

delta sets the resolution through the number of Chebyshev moments. Unlike in Dynamical correlators, it is not calibrated as a broadening, so compare distributions computed with the same delta. get_distribution also works with mode="ED". get_distribution_moments returns the raw Chebyshev moments of the rescaled operator and has no ED version. A maximum-entropy reconstruction from the moments \(\langle X^k\rangle\), distribution.get_distribution_maxent(sc, X=..., n=6), gives the smoothest distribution with the first n moments. With few moments it shows the envelope of the distribution, not its separate peaks.

Magnetization distribution of the Ising chain Magnetization distribution of the Ising chain

The figure shows \(P(m)\) for the magnetization per site \(m=\frac1n\sum_iS^z_i\) in the transverse-field Ising chain \(H=-\sum_iS^z_iS^z_{i+1}-h\sum_iS^x_i\), which is critical at \(h=1/2\). A small field \(10^{-4}\sum_iS^z_i\) selects one of the two ordered states. In the ordered phase (\(h=0.2\)), \(P(m)\) is one sharp peak close to the exact bulk value \(m=0.489\). At the critical field it is broad and flat, and in the paramagnet (\(h=1\)) it is one peak at \(m=0\). On the left, with \(n=12\), the DMRG curves lie on the ED curves (dashed lines). There the peaks are separated, because \(m\) only takes the values \(k/n\). On the right, with \(n=24\), these values are closer than the resolution set by delta, so the distribution looks smooth.

Code of this figure (docs/figures/correlators_distribution/)
compute.py
# Distribution P(m) of the magnetization per site m in the ground state of the
# transverse-field Ising chain H = -sum SzSz - h sum Sx - hz sum Sz, critical
# at h=1/2, from a Chebyshev (KPM) expansion; ED checks DMRG on the small chain.
import numpy as np
from dmrgpy import spinchain

hs = [0.2, 0.5, 1.0]
hz = 1e-4  # tiny longitudinal field that picks one of the two ordered states
n_small, n_large = 12, 24
scale, npol = 0.625, 96  # m in [-1/2,1/2] maps to [-0.8,0.8], 96 moments
# The same npol moments on both solvers: DMRG keeps int(3*scale/delta)+2 and
# ED 4*int(scale/delta), and half a moment more survives the int() rounding.
delta = 3*scale/(npol - 2 + 0.5)
delta_ed = scale/(npol/4 + 0.5)


def chain(n, h):
    """The chain and its magnetization per site m."""
    sc = spinchain.Spin_Chain(["S=1/2"]*n)
    sc.set_hamiltonian(-sum(sc.Sz[i]*sc.Sz[i+1] for i in range(n-1))
                       - h*sum(sc.Sx[i] for i in range(n))
                       - hz*sum(sc.Sz[i] for i in range(n)))
    sc.maxm, sc.kpmmaxm, sc.nsweeps = 60, 60, 12
    return sc, sum(sc.Sz[i] for i in range(n))*(1.0/n)


p_small, p_small_ed, p_large = [], [], []
for h in hs:
    sc, m = chain(n_small, h)
    m_small, p = sc.get_distribution(X=m, delta=delta, scale=scale)
    p_ed = sc.get_distribution(X=m, delta=delta_ed, scale=scale, xs=m_small,
                               mode="ED")[1]
    p_small.append(p.real)
    p_small_ed.append(p_ed.real)
    sc, m = chain(n_large, h)
    m_large, p = sc.get_distribution(X=m, delta=delta, scale=scale)
    p_large.append(p.real)

np.savez("data.npz", h=hs, n_small=n_small, n_large=n_large, m_small=m_small,
         p_small=p_small, p_small_ed=p_small_ed, m_large=m_large,
         p_large=p_large)
plot.py
import numpy as np
import style

d = np.load("data.npz")
h = d["h"]
names = {0.2: "ordered", 0.5: "critical", 1.0: "paramagnet"}
# bulk magnetization per site of the ordered phase (Pfeuty), m = (1-(2h)^2)^(1/8)/2
m_pfeuty = 0.5*(1-(2*h[0])**2)**0.125


def draw(theme):
    c = style.colors(theme)
    fig, (a, b) = style.figure(2)
    for k, hk in enumerate(h):
        lab = r"$h=%g$, %s" % (hk, names.get(float(hk), ""))
        a.plot(d["m_small"], d["p_small"][k], color=c["series"][k], label=lab)
        b.plot(d["m_large"], d["p_large"][k], color=c["series"][k], label=lab)
    for k in range(len(h)):
        a.plot(d["m_small"], d["p_small_ed"][k], **style.reference(theme),
               label="ED" if k == 0 else None)
    b.axvline(m_pfeuty, color=c["ink2"], lw=0.8, ls=":")
    b.text(m_pfeuty, b.get_ylim()[1]*0.97, " bulk $m$\n (exact)",
           color=c["ink2"], ha="right", va="top", fontsize=8)
    for ax, n in ((a, d["n_small"]), (b, d["n_large"])):
        ax.set_xlabel(r"magnetization per site $m$")
        ax.set_xlim(-0.55, 0.55)
        ax.set_ylim(bottom=0)
        ax.text(0.02, 0.97, r"$n=%d$" % n, transform=ax.transAxes,
                ha="left", va="top")
    a.set_ylabel(r"$P(m)$")
    a.legend(loc="upper left", bbox_to_anchor=(0.0, 0.9))
    return fig


style.main(draw, "correlators_distribution")

The many-body density of states, \(\rho(E)=\sum_n\delta(E-E_n)\), is not a distribution of this kind: get_distribution computes a ground-state expectation value, not a trace over all the states, and there is no method for \(\rho(E)\). For a small chain the full ED spectrum from get_excited(mode="ED", n=...) (Excited states) gives it directly.