Qutip

QuTiP 5

Scope

Use QuTiP for finite-dimensional quantum mechanics, quantum optics, Lindblad dynamics, trajectories, weak-coupling Bloch-Redfield models, and specialized Floquet, HEOM, and permutational-invariance methods. It is not a hardware execution SDK. Circuit and control functionality moved to separate QuTiP family packages.

This skill targets QuTiP 5.3.1, released 2026-08-04. QuTiP 5.3 requires Python 3.11 or newer. Its required distributions are NumPy (>=1.23.2), SciPy (>=1.9.2, excluding 1.16.0 and 1.17.0), and packaging.

Reproducible uv snapshot

Create a dedicated environment and pin every direct distribution:

uv venv --python 3.11
uv pip install "qutip==5.3.1"

For plots:

uv pip install "qutip[graphics]==5.3.1"

Optional QuTiP family packages are independently versioned:

uv pip install "qutip-qip==0.4.2"
uv pip install "qutip-qtrl==0.2.0"
uv pip install "qutip-jax==0.1.1"

• qutip-qip 0.4.2 (2026-06-23) is the production/stable circuit, gate, and noisy-device simulation package. Import from qutip_qip, not qutip.qip.
• qutip-qtrl 0.2.0 (2026-06-23) provides GRAPE and CRAB quantum optimal control. It is not a trajectory viewer. Import from qutip_qtrl, not qutip.control; PyPI still classifies it pre-alpha.
• qutip-jax 0.1.1 (2025-05-29) is the official JAX data backend for GPU and automatic-differentiation experiments. It is explicitly pre-alpha.
• qutip-cupy is an official QuTiP-organization repository, but it has no PyPI release and its own README says it is not officially released. Do not put an unreleased Git install into a reproducible workflow.

Use a project lockfile or a hash-generating uv pip compile workflow when transitive dependency identity must also be frozen.

Core tests cover small native CPU systems with known solutions. Optional JAX checks cover CPU conversion/evolution and differentiation only; GPU/MPI, large HEOM hierarchies, and experimentally realistic physics are unvalidated. Snippets with placeholders such as H, rho0, or c_ops are illustrative fragments to adapt after defining a consistent model.

Non-negotiable model contract

Before solving, record:

1. Units and convention. QuTiP equations normally set (\hbar=1). Hamiltonian entries are angular frequencies and rates have reciprocal-time units. A decay rate specified as 1/T1 needs no 2*pi factor. Convert cyclic frequency with (2\pi f); never mix Hz and rad/s.
2. Subsystem order. tensor(A, B, C) fixes subsystem indices 0, 1, 2. Preserve that order in every state, operator, collapse channel, and partial trace. obj.ptrace([0, 2]) keeps those subsystems; it does not trace them.
3. State validity. Check ket norm or density-matrix Hermiticity, unit trace, and eigenvalues above a stated negative tolerance. Tiny negative values may be numerical; material negativity invalidates a claimed state.
4. Generator meaning. A Lindblad channel with rate gamma is represented by sqrt(gamma) * A, not gamma * A. Define what each rate measures. For example, sqrt(gamma_phi / 2) * sigmaz() gives coherence decay exp(-gamma_phi * t).
5. Approximations. State rotating-wave, Born-Markov, secular, weak-coupling, bath-equilibrium, truncation, symmetry, and initial-factorization assumptions wherever used.
6. Numerics. Justify Hilbert truncation, output grid, integration method, tolerances, trajectory count, and random seeds. Report result.stats.
7. Convergence. Sweep every artificial cutoff: Fock dimension, time/frequency window and spacing, ODE tolerances, trajectories, Floquet harmonics, HEOM depth and bath exponents, or PIQS representation as applicable.

Qobj, dimensions, and tensor order

Prefer explicit imports and inspect both shape and structured dimensions:

from qutip import basis, qeye, sigmaz, tensor

psi = tensor(basis(2, 0), basis(3, 1))
z_on_first = tensor(sigmaz(), qeye(3))

assert psi.shape == (6, 1)
assert psi.dims == [[2, 3], [1]]
assert z_on_first.dims == [[2, 3], [2, 3]]
rho_first = psi.proj().ptrace(0)  # keep subsystem 0

Matrix shape alone is insufficient: two objects can both be 6-by-6 but encode different tensor factorizations. Read references/core_concepts.md before building composite, superoperator, or channel models.

Choose the solver by physics

Do not select a more specialized solver merely because it exists.

Deterministic open-system example

QuTiP 5.3 uses ordinary option dictionaries. e_ops, args, and options are keyword-only; the old mutable options object is gone.

import numpy as np
from qutip import basis, mesolve, sigmam, sigmaz

omega = 2.0
gamma = 0.15
tlist = np.linspace(0.0, 20.0, 401)
excited = basis(2, 0)

result = mesolve(
    0.5 * omega * sigmaz(),
    excited,
    tlist,
    c_ops=[np.sqrt(gamma) * sigmam()],
    e_ops={"sigma_z": sigmaz(), "excited": excited.proj()},
    options={
        "method": "adams",
        "atol": 1e-10,
        "rtol": 1e-8,
        "store_final_state": True,
        "normalize_output": False,
        "progress_bar": "",
    },
)

population = np.asarray(result.e_data["excited"])
assert np.max(np.abs(population - np.exp(-gamma * tlist))) < 2e-6
assert isinstance(result.stats, dict)

If the problem is stiff, compare bdf or lsoda; do not change an integrator without rerunning tolerance and invariant checks. QuTiP 5.3 also supports options={"matrix_form": True} in mesolve; benchmark and validate it before using it as a default.

Time-dependent systems

Prefer trusted Pythonic callables or numeric coefficient arrays. Do not create coefficient source strings from user input.

import numpy as np
from qutip import QobjEvo, sigmax, sigmaz

def envelope(t, amplitude, center, width):
    return amplitude * np.exp(-0.5 * ((t - center) / width) ** 2)

H = QobjEvo(
    [0.5 * sigmaz(), [sigmax(), envelope]],
    args={"amplitude": 0.2, "center": 5.0, "width": 1.0},
)
instantaneous_H = H(5.0)
H.arguments(amplitude=0.1)

The older f(t, args) coefficient signature is deprecated in 5.3 and is scheduled for removal in 5.5. See references/time_evolution.md.

For narrow pulses, the output tlist is not the adaptive integrator’s internal step schedule. Bound the solver’s max_step below half the narrowest pulse width, then reduce it further to check convergence of the pulse response. Tight relative/absolute tolerances alone can still miss a pulse sampled only in an idle region. See the QuTiP solver options.

Trajectories and stochastic solvers

import numpy as np
from qutip import basis, mcsolve, sigmam, sigmaz

tlist = np.linspace(0.0, 10.0, 201)
result = mcsolve(
    0.5 * sigmaz(),
    basis(2, 0),
    tlist,
    [np.sqrt(0.2) * sigmam()],
    e_ops=[basis(2, 0).proj()],
    ntraj=400,
    seeds=20260723,
    options={"keep_runs_results": False, "progress_bar": ""},
)

Report ntraj, result.seeds, uncertainty or repeated-seed sensitivity, and whether individual runs were retained. Reuse seeds=previous_result.seeds only when paired trajectories are intentional. ssesolve and smesolve use the boolean heterodyne argument, not legacy integer noise codes.