qutip/qutip: QuTiP 5.0.0a1

This is a pre-release.

QuTiP 5 is a redesign of many of the core components of QuTiP (Qobj, QobjEvo, solvers) to make them more consistent and more flexible.

Qobj may now be stored in either sparse or dense representations, and the two may be mixed sensibly as needed. QobjEvo is now used consistently throughout QuTiP, and the implementation has been substantially cleaned up. A new Coefficient class is used to represent the time-dependent factors inside QobjEvo.

The solvers have been rewritten to work well with the new data layer and the concept of Integrators which solve ODEs has been introduced. In future, new data layers may provide their own Integrators specialized to their representation of the underlying data.

Much of the user-facing API of QuTiP remains familiar, but there have had to be many small breaking changes. If we can make changes to easy migrating code from QuTiP 4 to QuTiP 5, please let us know.

Any extensive list of changes follows.

QuTiP 5 has been a large effort by many people over the last three years.

In particular:

• Jake Lishman led the implementation of the new data layer and coefficients.
• Eric Giguère led the implementation of the new QobjEvo interface and solvers.
• Boxi Li led the updating of QuTiP's QIP support and the creation of qutip_qip.

Other members of the QuTiP Admin team have been heavily involved in reviewing, testing and designing QuTiP 5:

• Alexander Pitchford
• Asier Galicia
• Nathan Shammah
• Shahnawaz Ahmed
• Neill Lambert
• Simon Cross

Two Google Summer of Code contributors updated the tutorials and benchmarks to QuTiP 5:

• Christian Staufenbiel updated many of the tutorials (https://github.com/qutip/qutip-tutorials/).
• Xavier Sproken update the benchmarks (https://github.com/qutip/qutip-benchmark/).

Four experimental data layers backends were written either as part of Google Summer of Code or as separate projects. While these are still alpha quality, the helped significantly to test the data layer API:

• qutip-tensorflow: a TensorFlow backend by Asier Galicia (https://github.com/qutip/qutip-tensorflow)
• qutip-cupy: a CuPy GPU backend by Felipe Bivort Haiek (https://github.com/qutip/qutip-cupy)
• qutip-tensornetwork: a TensorNetwork backend by Asier Galicia (https://github.com/qutip/qutip-tensornetwork)
• ``qutip-jax```: a JAX backend by Eric Giguère (https://github.com/qutip/qutip-jax)

We have also had many other contributors, whose specific contributions are detailed below:

• Pieter Eendebak (updated the required SciPy to 1.4+, #1982).
• Pieter Eendebak (reduced import times by setting logger names, #1981)
• Xavier Sproken (included C header files in the source distribution, #1971)
• Christian Staufenbiel (added support for multiple collapse operators to the Floquet solver, #1962)
• Christian Staufenbiel (fixed the basis used in the Floquet Master Equation solver, #1952)
• Christian Staufenbiel (allowed the bloch_redfield_tensor function to accept strings and callables for a_ops, #1951)
• Henrique Silvéro (allowed qutip_qip to be imported as qutip.qip, #1920)
• Florian Hopfmueller (added a vastly improved implementations of process_fidelity and average_gate_fidelity, #1712, #1748, #1788)
• Felipe Bivort Haiek (fixed inaccuracy in docstring of the dense implementation of negation, #1608)
• Rajath Shetty (added support for specifying colors for individual points, vectors and states display by qutip.Bloch, #1335)

Previously Qobj data was stored in a SciPy-like sparse matrix. Now the representation is flexible. Implementations for dense and sparse formats are included in QuTiP and custom implementations are possible. QuTiP's performance on dense states and operators is significantly improved as a result.

Some highlights:

• The data is still acessible via the .data attribute, but is now an instance of the underlying data type instead of a SciPy-like sparse matrix. The operations available in qutip.core.data may be used on .data, regardless of the data type.
• Qobj with different data types may be mixed in arithmetic and other operations. A sensible output type will be automatically determined.
• The new .to(...) method may be used to convert a Qobj from one data type to another. E.g. .to("dense") will convert to the dense representation and .to("csr") will convert to the sparse type.
• Many Qobj methods and methods that create Qobj now accepted a dtype parameter that allows the data type of the returned Qobj to specified.
• The new & operator may be used to obtain the tensor product.
• The new @ operator may be used to obtain the matrix / operator product. bar @ ket returns a scalar.
• The new .contract() method will collapse 1D subspaces of the dimensions of the Qobj.
• The new .logm() method returns the matrix logarithm of an operator.
• The methods .set_data, .get_data, .extract_state, .eliminate_states, .evaluate and .check_isunitary have been removed.

The QobjEvo type for storing time-dependent quantum objects has been significantly expanded, standardized and extended. The time-dependent coefficients are now represented using a new Coefficient type that may be independently created and manipulated if required.

Some highlights:

• The .compile() method has been removed. Coefficients specified as strings are automatically compiled if possible and the compilation is cached across different Python runs and instances.
• Mixing coefficient types within a single Qobj is now supported.
• Many new attributes were added to QobjEvo for convenience. Examples include .dims, .shape, .superrep and .isconstant.
• Many old attributes such as .cte, .use_cython, .type, .const, and .coeff_file were removed.
• A new Spline coefficient supports spline interpolations of different orders. The old Cubic_Spline coefficient has been removed.
• The new .arguments(...) method allows additional arguments to the underlying coefficient functions to be updated.
• The _step_func_coeff argument has been replaced by the order parameter. _step_func_coeff=False is equivalent to order=3. _step_func_coeff=True is equivalent to order=0. Higher values of order gives spline interpolations of higher orders.

The solvers in QuTiP have been heavily reworked and standardized. Under the hood solvers now make use of swappable ODE Integrators. Many Integrators are included (see the list below) and custom implementations are possible. Solvers now consistently accept a QobjEvo instance at the Hamiltonian or Liouvillian, or any object which can be passed to the QobjEvo constructor.

A breakdown of highlights follows.

All solvers:

• Solver options are now supplied in an ordinary Python dict. qutip.Options is deprecated and returns a dict for backwards compatibility.
• A specific ODE integrator may be selected by supplying a method option.
• Each solver provides a class interface. Creating an instance of the class allows a solver to be run multiple times for the same system without having to repeatedly reconstruct the right-hand side of the ODE to be integrated.
• A QobjEvo instance is accepted for most operators, e.g., H, c_ops, e_ops, a_ops.
• The progress bar is now selected using the progress_bar option. A new progess bar using the tqdm Python library is provided.
• Dynamic arguments, where the value of an operator depends on the current state of the evolution, have been removed. They may be re-implemented later if there is demand for them.

Integrators:

• The SciPy zvode integrator is available with the BDF and Adams methods as bdf and adams.
• The SciPy dop853 integrator (an eighth order Runge-Kutta method by Dormand & Prince) is available as dop853.
• The SciPy lsoda integrator is available as lsoda.
• QuTiP's own implementation of Verner's "most efficient" Runge-Kutta methods of order 7 and 9 are available as vern7 and vern9. See http://people.math.sfu.ca/~jverner/ for a description of the methods.
• QuTiP's own implementation of a solver that directly diagonalizes the the system to be integrated is available as diag. It only works on time-independent systems and is slow to setup, but once the diagonalization is complete, it generates solutions very quickly.
• QuTiP's own implementatoin of an approximate Krylov subspace integrator is available as krylov. This integrator is only usable with sesolve.

Result class:

• A new .e_data attribute provides expectation values as a dictionary. Unlike .expect, the values are provided in a Python list rather than a numpy array, which better supports non-numeric types.
• The contents of the .stats attribute changed significantly and is now more consistent across solvers.

Monte-Carlo Solver (mcsolve):

• The system, H, may now be a super-operator.
• The seed parameter now supports supplying numpy SeedSequence or Generator types.
• The new timeout and target_tol parameters allow the solver to exit early if a timeout or target tolerance is reached.
• The ntraj option no longer supports a list of numbers of trajectories. Instead, just run the solver multiple times and use the class MCSolver if setting up the solver uses a significant amount of time.
• The map_func parameter has been replaced by the map option. In addition to the existing serial and parallel values, the value loky may be supplied to use the loky package to parallelize trajectories.
• The result returned by mcsolve now supports calculating photocurrents and calculating the steady state over N trajectories.
• The old parfor parallel execution function has been removed from qutip.parallel. Use parallel_map or loky_map instead.

Bloch-Redfield Master Equation Solver (brmesolve):

• The a_ops and spectra support implementaitons been heavily reworked to reuse the techniques from the new Coefficient and QobjEvo classes.
• The use_secular parameter has been removed. Use sec_cutoff=-1 instead.
• The required tolerance is now read from qutip.settings.

Krylov Subspace Solver (krylovsolve):

• The Krylov solver is now implemented using SESolver and the krylov ODE integrator. The function krylovsolve is maintained for convenience and now supports many more options.
• The sparse parameter has been removed. Supply a sparse Qobj for the Hamiltonian instead.

Floquet Solver (fsesolve and fmmesolve):

• The Floquet solver has been rewritten to use a new FloquetBasis class which manages the transformations from lab to Floquet basis and back.
• Many of the internal methods used by the old Floquet solvers have been removed. The Floquet tensor may still be retried using the function floquet_tensor.
• The Floquet Markov Master Equation solver has had many changes and new options added. The environment temperature may be specified using w_th, and the result states are stored in the lab basis and optionally in the Floquet basis using store_floquet_state.
• The spectra functions supplied to fmmesolve must now be vectorized (i.e. accept and return numpy arrays for frequencies and densities) and must accept negative frequence (i.e. usually include a w > 0 factor so that the returned densities are zero for negative frequencies).
• The number of sidebands to keep, kmax may only be supplied when using the FMESolver
• The Tsteps parameter has been removed from both fsesolve and fmmesolve. The precompute option to FloquetBasis may be used instead.

Evolution of State Solver (essovle):

• The function essolve has been removed. Use the diag integration method with sesolve or mesolve instead.

Steady-state solvers (steadystate module):

• The method parameter and solver parameters have been separated. Previously they were mixed together in the method parameter.
• The previous options are now passed as parameters to the steady state solver and mostly passed through to the underlying SciPy functions.
• The logging and statistics have been removed.

Correlation functions (correlation module):

• A new correlation_3op function has been added. It supports MESolver or BRMESolver.
• The correlation, correlation_4op, and correlation_ss functions have been removed.
• Support for calculating correlation with mcsolve has been removed.

Propagators (propagator module):

• A class interface, qutip.Propagator, has been added for propagators.
• Propagation of time-dependent systems is now supported using QobjEvo.
• The unitary_mode and parallel options have been removed.

Correlation spectra (spectrum module):

• The functions spectrum_ss and spectrum_pi have been removed and are now internal functions.
• The use_pinv parameter for spectrum has been removed and the functionality merged into the solver parameter. Use solver="pi" instead.

There have been numerous other small changes to core QuTiP features:

• qft(...) the function that returns the quantum Fourier transform operator was moved from qutip.qip.algorithm into qutip.
• The Bloch-Redfield solver tensor, brtensor, has been moved into qutip.core. See the section above on the Bloch-Redfield solver for details.
• The functions mat2vec and vec2mat for transforming states to and from super-operator states have been renamed to stack_columns and unstack_columns.
• The function liouvillian_ref has been removed. Used liouvillian instead.
• The superoperator transforms super_to_choi, choi_to_super, choi_to_kraus, choi_to_chi and chi_to_choi have been removed. Used to_choi, to_super, to_kraus and to_chi instead.
• All of the random object creation functions now accepted a numpy Generator as a seed.
• The dims parameter of all random object creation functions has been removed. Supply the dimensions as the first parameter if explicit dimensions are required.
• The function rand_unitary_haar has been removed. Use rand_unitary(distribution="haar") instead.
• The functions rand_dm_hs and rand_dm_ginibre have been removed. Use rand_dm(distribution="hs") and rand_dm(distribution="ginibre") instead.
• The function rand_ket_haar has been removed. Use rand_ket(distribution="haar") instead.
• The measurement functions have had the target parameter for expanding the measurement operator removed. Used expand_operator to expand the operator instead.
• qutip.Bloch now supports applying colours per-point, state or vector in add_point, add_states, and add_vectors.

Previously qutip.settings was an ordinary module. Now qutip.settings is an instance of a settings class. All the runtime modifiable settings for core operations are in qutip.settings.core. The other settings are not modifiable at runtime.

• Removed load. reset and save functions.
• Removed .debug, .fortran, .openmp_thresh.
• New .compile stores the compilation options for compiled coefficients.
• New .core["rtol"] core option gives the default relative tolerance used by QuTiP.
• The absolute tolerance setting .atol has been moved to .core["atol"].

• qutip.qip has been moved into its own package, qutip-qip. Once installed, qutip-qip is available as either qutip.qip or qutip_qip. Some widely useful gates have been retained in qutip.gates.
• qutip.lattice has been moved into its own package, qutip-lattice. It is available from <https://github.com/qutip/qutip-lattice>.
• qutip.sparse has been removed. It contained the old sparse matrix representation and is replaced by the new implementation in qutip.data.
• qutip.piqs functions are no longer available from the qutip namespace. They are accessible from qutip.piqs instead.

• Support has been added for 64-bit integer sparse matrix indices, allowing sparse matrices with up to 2**63 rows and columns. This support needs to be enabled at compilation time by calling setup.py and passing --with-idxint-64.

• Support for OpenMP has been removed. If there is enough demand and a good plan for how to organize it, OpenMP support may return in a future QuTiP release.
• The qutip.parfor function has been removed. Use qutip.parallel_map instead.
• qutip.graph has been removed and replaced by SciPy's graph functions.
• qutip.topology has been removed. It contained only one function berry_curvature.
• The ~/.qutip/qutiprc config file is no longer supported. It contained settings for the OpenMP support.