QuTiP 5.0.0
QuTiP 5.0.0
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.
A notebook to help with migration is available on colab.
An extensive list of changes follows.
Contributors
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
- Paul Menczel
Two Google Summer of Code contributors updated the tutorials and benchmarks to
QuTiP 5:
- Christian Staufenbiel updated many of the tutorials.
- Xavier Sproken update the benchmarks.
During an internship at RIKEN, Patrick Hopf created a new quantum control method and
improved the existing methods interface:
- Patrick Hopf created new quantum control package.
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, they 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/)
Finally, Yuji Tamakoshi updated the visualization function and added animation
functions as part of Google Summer of Code project.
We have also had many other contributors, whose specific contributions are
detailed below:
- Pieter Eendebak (updated the required SciPy to 1.5+, #1982)
- Pieter Eendebak (reduced import times by setting logger names, #1981)
- Pieter Eendebak (Allow scipy 1.12 to be used with qutip, #2354)
- 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 fora_ops
, #1951) - Christian Staufenbiel (Add a guide on Superoperators, Pauli Basis and Channel Contraction, #1984)
- Henrique Silvéro (allowed
qutip_qip
to be imported asqutip.qip
, #1920) - Florian Hopfmueller (added a vastly improved implementations of
process_fidelity
andaverage_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) - Rochisha Agarwal (Add dtype to printed ouput of qobj, #2352)
- Kosuke Mizuno (Add arguments of plot_wigner() and plot_wigner_fock_distribution() to specify parameters for wigner(), #2057)
- Matt Ord (Only pre-compute density matrices if keep_runs_results is False, #2303)
- Daniel Moreno Galán (Add the possibility to customize point colors as in V4 and fix point plot behavior for 'l' style, #2303)
- Sola85 (Fixed simdiag not returning orthonormal eigenvectors, #2269)
- Edward Thomas (Fix LaTeX display of Qobj state in Jupyter cell outputs, #2272)
- Bogdan Reznychenko (Rework
kraus_to_choi
making it faster, #2284) - gabbence95 (Fix typos in
expect
documentation, #2331) - lklivingstone (Added repr to QobjEvo, #2111)
- Yuji Tamakoshi (Improve print(qutip.settings) by make it shorter, #2113)
- khnikhil (Added fermionic annihilation and creation operators, #2166)
- Daniel Weiss (Improved sampling algorithm for mcsolve, #2218)
- SJUW (Increase missing colorbar padding for matrix_histogram_complex() from 0 to 0.05, #2181)
- Valan Baptist Mathuranayagam (Changed qutip-notebooks to qutip-tutorials and fixed the typo in the link redirecting to the changelog section in the PR template, #2107)
- Gerardo Jose Suarez (Added information on sec_cutoff to the documentation, #2136)
- Cristian Emiliano Godinez Ramirez (Added inherited members to API doc of MESolver, SMESolver, SSESolver, NonMarkovianMCSolver, #2167)
- Andrey Rakhubovsky (Corrected grammar in Bloch-Redfield master equation documentation, #2174)
- Rushiraj Gadhvi (qutip.ipynbtools.version_table() can now be called without Cython installed, #2110)
- Harsh Khilawala (Moved HTMLProgressBar from qutip/ipynbtools.py to qutip/ui/progressbar.py, #2112)
- Avatar Srinidhi P V (Added new argument bc_type to take boundary conditions when creating QobjEvo, #2114)
- Andrey Rakhubovsky (Fix types in docstring of projection(), #2363)
Qobj changes
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 inqutip.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 aQobj
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 createQobj
now accepted adtype
parameter that allows the data type of the returnedQobj
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
theQobj
. - 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 property
dtype
return the representation of the data used. - The new
data_as
allow to obtain the data as a common python formats:
numpy array, scipy sparse matrix, JAX Array, etc.
QobjEvo changes
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 oldCubic_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 theorder
parameter._step_func_coeff=False
is equivalent toorder=3
.
_step_func_coeff=True
is equivalent toorder=0
. Higher values
oforder
gives spline interpolations of higher orders. - The spline type can take
bc_type
to control the boundary conditions. - QobjEvo can be creating from the multiplication of a Qobj with a coefficient:
oper * qutip.coefficient(f, args=args)
is equivalent to
qutip.QobjEvo([[oper, f]], args=args)
. - Coefficient function can be defined in a pythonic manner:
def f(t, A, w)
.
The dictionaryargs
second argument is no longer needed.
Function using the exactf(t, args)
signature will use the old method for
backward compatibility.
Solver changes
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 thetqdm
Python library is provided. - Dynamic arguments, where the value of an operator depends on
the current state of the evolution interface reworked. Now a property of the
solver is to be used as an arguments:
args={"state": MESolver.StateFeedback(default=rho0)}
Integrators:
- The SciPy zvode integrator is available with the BDF and
Adams methods asbdf
andadams
. - The SciPy dop853 integrator (an eighth order Runge-Kutta method by
Dormand & Prince) is available asdop853
. - 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 asvern7
andvern9
. 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 asdiag
. 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 askrylov
. This integrator is only usable withsesolve
.
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 numpySeedSequence
or
Generator
types. - The new
timeout
andtarget_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 classMCSolver
if setting up the solver uses a significant amount of time. - The
map_func
parameter has been replaced by themap
option. - A loky based parallel map as been added.
- A mpi based parallel map as been added.
- 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
. Useparallel_map
,loky_map
ormpi_pmap
instead. - Added improved sampling options which converge much faster when the
probability of collapse is small.
Non Markovian Monte-Carlo Solver (nm_mcsolve):
- New Monte-Carlo Solver supporting negative decay rates.
- Based on the influence martingale approach, Donvil et al., Nat Commun 13, 4140 (2022).
- Most of the improvements made to the regular Monte-Carlo solver are also available here.
- The value of the influence martingale is available through the
.trace
attribute of the result.
Stochastic Equation Solvers (ssesolve, smesolve)
- Function call greatly changed: many keyword arguments are now options.
- m_ops and dW_factors are now changed from the default from the new class interface only.
- Use the same parallel maps as mcsolve: support for loky and mpi map added.
- End conditions
timeout
andtarget_tol
added. - The
seed
parameter now supports supplying numpySeedSequence
. - Wiener function is now available as a feedback.
Bloch-Redfield Master Equation Solver (brmesolve):
- The
a_ops
andspectra
support implementations been heavily reworked to
reuse the techniques from the new Coefficient and QobjEvo classes. - The
use_secular
parameter has been removed. Usesec_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 thekrylov
ODE integrator. The functionkrylovsolve
is maintained for convenience
and now supports many more options. - The
sparse
parameter has been removed. Supply a sparseQobj
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 functionfloquet_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 usingstore_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 aw > 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
theFMESolver
- The
Tsteps
parameter has been removed from bothfsesolve
and
fmmesolve
. Theprecompute
option toFloquetBasis
may be used
instead.
Evolution of State Solver (essovle):
- The function
essolve
has been removed. Use thediag
integration
method withsesolve
ormesolve
instead.
Steady-state solvers (steadystate module):
- The
method
parameter andsolver
parameters have been separated. Previously
they were mixed together in themethod
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 supportsMESolver
orBRMESolver
. - The
correlation
,correlation_4op
, andcorrelation_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
andparallel
options have been removed.
Correlation spectra (spectrum module):
- The functions
spectrum_ss
andspectrum_pi
have been removed and
are now internal functions. - The
use_pinv
parameter forspectrum
has been removed and the
functionality merged into thesolver
parameter. Usesolver="pi"
instead.
Hierarchical Equation of Motion Solver (HEOM)
- Updated the solver to use the new QuTiP integrators and data layer.
- Updated all the HEOM tutorials to QuTiP 5.
- Added support for combining bosonic and fermionic baths.
- Sped up the construction of the RHS of the HEOM solver by a factor of 4x.
- As in QuTiP 4, the HEOM supports arbitrary spectral densities, bosonic and fermionic baths, Páde and Matsubara expansions of the correlation functions, calculating the Matsubara terminator and inspection of the ADOs (auxiliary density operators).
QuTiP core
There have been numerous other small changes to core QuTiP features:
qft(...)
the function that returns the quantum Fourier
transform operator was moved fromqutip.qip.algorithm
intoqutip
.- 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
andvec2mat
for transforming states to and
from super-operator states have been renamed tostack_columns
and
unstack_columns
. - The function
liouvillian_ref
has been removed. Usedliouvillian
instead. - The superoperator transforms
super_to_choi
,choi_to_super
,
choi_to_kraus
,choi_to_chi
andchi_to_choi
have been removed.
Usedto_choi
,to_super
,to_kraus
andto_chi
instead. - All of the random object creation functions now accepted a
numpyGenerator
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
andrand_dm_ginibre
have been removed.
Userand_dm(distribution="hs")
andrand_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. Usedexpand_operator
to expand the operator instead. qutip.Bloch
now supports applying colours per-point, state or vector in
add_point
,add_states
, andadd_vectors
.- Dimensions use a class instead of layered lists.
- Allow measurement functions to support degenerate operators.
- Add
qeye_like
andqzero_like
. - Added fermionic annihilation and creation operators.
QuTiP settings
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
andsave
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"]
.
Visualization
- Added arguments to
plot_wigner
andplot_wigner_fock_distribution
to specify parameters forwigner
. - Removed
Bloch3D
. The same functionality is provided byBloch
. - Added
fig
,ax
andcmap
keyword arguments to all visualization functions. - Most visualization functions now respect the
colorblind_safe
setting. - Added new functions to create animations from a list of
Qobj
or directly from solver results with saved states.
Package reorganization
qutip.qip
has been moved into its own package, qutip-qip. Once installed, qutip-qip is available as eitherqutip.qip
orqutip_qip
. Some widely useful gates have been retained inqutip.gates
.qutip.control
has been moved to qutip-qtrl and once installed qutip-qtrl is available as eitherqutip.control
orqutip_qtrl
. Note thatquitp_qtrl
is provided primarily for backwards compatibility. Improvements to optimal control will take place in the newqutip_qoc
package.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 inqutip.data
.qutip.piqs
functions are no longer available from thequtip
namespace. They are accessible fromqutip.piqs
instead.
Miscellaneous
- 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 callingsetup.py
and passing
--with-idxint-64
.
Feature removals
- 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. Usequtip.parallel_map
instead. qutip.graph
has been removed and replaced by SciPy's graph functions.qutip.topology
has been removed. It contained only one functionberry_curvature
.- The
~/.qutip/qutiprc
config file is no longer supported. It contained settings for the OpenMP support. - Deprecate
three_level_atom
- Deprecate
orbital
Changes from QuTiP 5.0.0b1:
Features
- Add dtype to printed ouput of qobj (#2352 by Rochisha Agarwal)
Miscellaneous
- Allow scipy 1.12 to be used with qutip. (#2354 by Pieter Eendebak)