Skip to content

Open systems

On this page we compute the steady state and the slowest relaxation modes of a chain coupled to an environment, described by a Lindblad master equation,

\[\frac{d\rho}{dt}=\mathcal L\rho=-i[H,\rho]+\sum_\mu\Big(L_\mu\rho L_\mu^\dagger -\tfrac12\{L_\mu^\dagger L_\mu,\rho\}\Big).\]

The jump operators \(L_\mu\) describe what the environment does: decay, pumping, dephasing, dissipative hopping.

Steady state and dissipative gap

\(\mathcal L\) is linear in \(\rho\), so it has eigenvalues \(\lambda_k\) with \(\mathrm{Re}\,\lambda_k\le 0\). The steady state is the eigenvector with \(\lambda=0\). The slowest mode sets the relaxation time through the dissipative gap \(\Delta=-\mathrm{Re}\,\lambda_1\), meaning that a generic state approaches the steady state as \(e^{-\Delta t}\). The way we do this is by writing \(\rho\) as a matrix product state on a chain of twice the length, one site for the ket and one for the bra index of each physical site. Then \(\mathcal L\) becomes an operator on that chain, and we take its slowest eigenvalues from time evolution, as in Excited states.

As a reference, let us take an XXZ chain pumped up at the left end, pumped down at the right end, and dephased on every site:

from dmrgpy import spinchain
from dmrgpy.lindblad import Lindblad_Chain
n = 6
sc = spinchain.Spin_Chain(["S=1/2" for i in range(n)], itensor_version="python")
h = sum(sc.Sx[i]*sc.Sx[i+1] + sc.Sy[i]*sc.Sy[i+1] + 0.5*sc.Sz[i]*sc.Sz[i+1]
        for i in range(n-1))
sc.set_hamiltonian(h)
sp = sc.Sx[0] + 1j*sc.Sy[0]          # pumps spin up at the left end
sm = sc.Sx[n-1] - 1j*sc.Sy[n-1]      # pumps spin down at the right end
jumps = [0.6*sp, 0.6*sm] + [0.4*sc.Sz[i] for i in range(n)]
lc = Lindblad_Chain(sc, jumps)
lam, res, obs = lc.get_spectrum()           # lam[0] ~ 0 is the steady state
gap = -lam[1].real                          # dissipative gap
mz = [lc.steady_state_vev(sc.Sz[i]) for i in range(n)]

lam is sorted by real part, the steady state first. res[k] is how much \(\lambda_k\) moves when the last snapshot is left out; only the eigenvalues that do not move are converged. mz is the magnetization profile of the steady state, a linear ramp from \(+0.18\) to \(-0.18\) that carries a spin current through the chain. Here the gap is \(0.0927\) against \(0.0926\) from ED, and the profile agrees with ED to \(3\times10^{-4}\). The keywords are those of get_krylov_spectrum, with one difference:

  • dt: the time step (default 2.0). The evolution must run for several times \(1/\Delta\), so nk*dt has to be large when the gap is small.
  • substeps: the number of time steps per snapshot to start from (default 1). A strongly dissipative \(\mathcal L\) makes long steps unstable, so a step that does not conserve the trace is redone with twice as many. Only the cost changes, not dt.
  • mode="ED": the dense \(\mathcal L\), as a reference on a small chain.

When \(\mathcal L\) conserves a quantity, for example the particle number for number-conserving jumps, every sector has its own steady state, and several eigenvalues come back at \(\lambda\approx 0\). To stay in one sector we start from product states of that sector:

  • start=(psi, phi): one local basis state per site (the occupation, for bosons). The evolution starts from \(|\psi\rangle\langle\psi|\) and from the traceless \(|\psi\rangle\langle\phi|\), and the sector of \(\psi\) is kept, for example start=([1,1,0],[0,1,1]) for two bosons on three sites. Modes of other sectors, which enter through small numerical errors, are left out of lam.

Backends

Lindblad_Chain works for spin and boson chains on itensor_version="python" and 3. Fermions are not supported.