Release notes

Release 0.5.0 (development release)

New features

  • Optimization callback functionalities has been improved. A dedicated Callback class is added which is able to access the optimizer, the cost function, the parameters as well as gradients, during the optimization. In addition, multiple callbacks can be specified. This opens up the endless possiblities of customizing the the optimization progress with schedulers, trackers, heuristics, tricks, etc. (#219)

  • Tensorboard based optimization tracking is added as a builtin Callback class: TensorboardCallback. It can automatically track costs as well as all trainable parameters during optimization in realtime. Tensorboard can be most conveniently viewed from VScode. (#219)

    import numpy as np
    from import Optimizer, TensorboardCallback
    def cost_fn():
    def as_dB(cost):
        delta = np.sqrt(np.log(1 / (abs(cost) ** 2)) / (2 * np.pi))
        cost_dB = -10 * np.log10(delta**2)
        return cost_dB
    tb_cb = TensorboardCallback(cost_converter=as_dB, track_grads=True)
    opt = Optimizer(euclidean_lr = 0.001);
    opt.minimize(cost_fn, max_steps=200, by_optimizing=[...], callbacks=tb_cb)
    # Logs will be stored in `tb_cb.logdir` which defaults to `./tb_logdir/...` but can be customized.
    # VScode can be used to open the Tensorboard frontend for live monitoring.
    # Or, in command line: `tensorboard --logdir={tb_cb.logdir}` and open link in browser.

Breaking Changes

  • The previous callback argument to Optimizer.minimize is now callbacks since we can now pass multiple callbacks to it.


  • Fixed flaky optimization tests and removed tf dependency. (#224) (#233)

Bug fixes

  • Unpins package versions in that got mistakenly pinned in 0.4.0. (#223)

  • fixing a bug with the Dgate optimization (#232)



Filippo Miatto, Sebastian Duque Mesa

This release contains contributions from (in alphabetical order): Zeyue Niu

Release 0.4.0 (current release)

New features

  • Ray-based distributed trainer is now added to training.trainer. It acts as a replacement for for loops and enables the parallelization of running many circuits as well as their optimizations. To install the extra dependencies: pip install .[ray]. (#194)

    from mrmustard.lab import Vacuum, Dgate, Ggate
    from mrmustard.physics import fidelity
    from import map_trainer
    def make_circ(x=0.):
        return Ggate(num_modes=1, symplectic_trainable=True) >> Dgate(x=x, x_trainable=True, y_trainable=True)
    def cost_fn(circ=make_circ(0.1), y_targ=0.):
        target = Gaussian(1) >> Dgate(-1.5, y_targ)
        s = Vacuum(1) >> circ
        return -fidelity(s, target)
    # Use case 0: Calculate the cost of a randomly initialized circuit 5 times without optimizing it.
    results_0 = map_trainer(
    # Use case 1: Run circuit optimization 5 times on randomly initialized circuits.
    results_1 = map_trainer(
    # Use case 2: Run circuit optimization 2 times on randomly initialized circuits with custom parameters.
    results_2 = map_trainer(
            {'x': 0.1, 'euclidean_lr': 0.005, 'max_steps': 50, 'HBAR': 1.},
            {'x': -0.7, 'euclidean_lr': 0.1, 'max_steps': 2, 'HBAR': 2.},
  • Sampling for homodyne measurements is now integrated in Mr Mustard: when no measurement outcome value is specified by the user, a value is sampled from the reduced state probability distribution and the conditional state on the remaining modes is generated. (#143)

    import numpy as np
    from mrmustard.lab import Homodyne, TMSV, SqueezedVacuum
    # conditional state from measurement
    conditional_state = TMSV(r=0.5, phi=np.pi)[0, 1] >> Homodyne(quadrature_angle=np.pi/2)[1]
    # measurement outcome
    measurement_outcome = SqueezedVacuum(r=0.5) >> Homodyne()
  • The optimizer minimize method now accepts an optional callback function, which will be called at each step of the optimization and it will be passed the step number, the cost value, and the value of the trainable parameters. The result is added to the callback_history attribute of the optimizer. (#175)

  • the Math interface now supports linear system solving via math.solve. (#185)

  • We introduce the tensor wrapper MMTensor (available in math.mmtensor) that allows for a very easy handling of tensor contractions. Internally MrMustard performs lots of tensor contractions and this wrapper allows one to label each index of a tensor and perform contractions using the @ symbol as if it were a simple matrix multiplication (the indices with the same name get contracted). (#185)

    from mrmustard.math.mmtensor import MMTensor
    # define two tensors
    A = MMTensor(np.random.rand(2, 3, 4), axis_labels=["foo", "bar", "contract"])
    B = MMTensor(np.random.rand(4, 5, 6), axis_labels=["contract", "baz", "qux"])
    # perform a tensor contraction
    C = A @ B
    C.axis_labels  # ["foo", "bar", "baz", "qux"]
    C.shape # (2, 3, 5, 6)
    C.tensor # extract actual result
  • MrMustard’s settings object (accessible via from mrmustard import settings) now supports SEED (an int). This will give reproducible results whenever randomness is involved. The seed is assigned randomly by default, and it can be reassigned again by setting it to None: settings.SEED = None. If one desires, the seeded random number generator is accessible directly via settings.rng (e.g. settings.rng.normal()). (#183)

  • The Circuit class now has an ascii representation, which can be accessed via the repr method. It looks great in Jupyter notebooks! There is a new option at settings.CIRCUIT_DECIMALS which controls the number of decimals shown in the ascii representation of the gate parameters. If None, only the name of the gate is shown. (#196)

  • PNR sampling from Gaussian circuits using density matrices can now be performed faster. When all modes are detected, this is done by replacing math.hermite_renormalized by math.hermite_renormalized_diagonal. If all but the first mode are detected, math.hermite_renormalized_1leftoverMode can be used. The complexity of these new methods is equal to performing a pure state simulation. The methods are differentiable, so that they can be used for defining a cost function. (#154)

  • MrMustard repo now provides a fully furnished vscode development container and a Dockerfile. To find out how to use dev containers for development check the documentation here. (#214)

Breaking changes


  • The Dgate is now implemented directly in MrMustard (instead of on The Walrus) to calculate the unitary and gradients of the displacement gate in Fock representation, providing better numerical stability for larger cutoff and displacement values. (#147) (#211)

  • Now the Wigner function is implemented in its own module and uses numba for speed. (#171)

    from mrmustard.utils.wigner import wigner_discretized
    W, Q, P = wigner_discretized(dm, q, p) # dm is a density matrix
  • Calculate marginals independently from the Wigner function thus ensuring that the marginals are physical even though the Wigner function might not contain all the features of the state within the defined window. Also, expose some plot parameters and return the figure and axes. (#179)

  • Allows for full cutoff specification (index-wise rather than mode-wise) for subclasses of Transformation. This allows for a more compact Fock representation where needed. (#181)

  • The mrmustard.physics.fock module now provides convenience functions for applying kraus operators and choi operators to kets and density matrices. (#180)

    from mrmustard.physics.fock import apply_kraus_to_ket, apply_kraus_to_dm, apply_choi_to_ket, apply_choi_to_dm
    ket_out = apply_kraus_to_ket(kraus, ket_in, indices)
    dm_out = apply_choi_to_dm(choi, dm_in, indices)
    dm_out = apply_kraus_to_dm(kraus, dm_in, indices)
    dm_out = apply_choi_to_ket(choi, ket_in, indices)
  • Replaced norm with probability in the repr of State. This improves consistency over the old behaviour (norm was the sqrt of prob if the state was pure and prob if the state was mixed). (#182)

  • Added two new modules (physics.bargmann and physics.husimi) to host the functions related to those representations, which have been refactored and moved out of physics.fock. (#185)

  • The internal type system in MrMustard has been beefed up with much clearer types, like ComplexVector, RealMatrix, etc… as well as a generic type Batch, which can be parametrized using the other types, like Batch[ComplexTensor]. This will allow for better type checking and better error messages. (#199)

  • Added multiple tests and improved the use of Hypothesis. (#191)

  • The fock.autocutoff function now uses the new diagonal methods for calculating a probability-based cutoff. Use settings.AUTOCUTOFF_PROBABILITY to set the probability threshold. (#203)

  • The unitary group optimization (for the interferometer) and the orthogonal group optimization (for the real interferometer) have been added. The symplectic matrix that describes an interferometer belongs to the intersection of the orthogonal group and the symplectic group, which is a unitary group, so we needed both. (#208)

Bug fixes

  • The Dgate and the Rgate now correctly parse the case when a single scalar is intended as the same parameter of a number of gates in parallel. (#180)

  • The trace function in the fock module was giving incorrect results when called with certain choices of modes. This is now fixed. (#180)

  • The purity function for fock states no longer normalizes the density matrix before computing the purity. (#180)

  • The function dm_to_ket no longer normalizes the density matrix before diagonalizing it. (#180)

  • The internal fock representation of states returns the correct cutoffs in all cases (solves an issue when a pure dm was converted to ket). (#184)

  • The ray related tests were hanging in github action causing tests to halt and fail. Now ray is forced to init with 1 cpu when running tests preventing the issue. (#201)

  • Various minor bug fixes. (#202)

  • Fixed the issue that the optimization of the interferometer was using orthogonal group optimization rather than unitary. (#208)

  • Fixes a slicing issue that arises when we compute the fidelity between gaussian and fock states. (#210)

  • The sign of parameters in the circuit drawer are now displayed correctly. (#209)

  • Fixed a bug in the Gaussian state which caused its covariance matrix to be multiplied by hbar/2 twice. Adds the argument modes to Ggate. (#212)

  • Fixes a bug in the cutoffs of the choi operator. (#216)



This release contains contributions from (in alphabetical order): Robbe De Prins, Sebastian Duque Mesa, Filippo Miatto, Zeyue Niu, Yuan Yao

Release 0.3.0

New features

  • Can switch progress bar on and off (default is on) from the settings via settings.PROGRESSBAR = True/False. (#128)

  • States in Gaussian and Fock representation now can be concatenated.

    from mrmustard.lab.states import Gaussian, Fock
    from mrmustard.lab.gates import Attenuator
    # concatenate pure states
    fock_state = Fock(4)
    gaussian_state = Gaussian(1)
    pure_state = fock_state & gaussian_state
    # also can concatenate mixed states
    mixed1 = fock_state >> Attenuator(0.8)
    mixed2 = gaussian_state >> Attenuator(0.5)
    mixed_state = mixed1 & mixed2


  • Parameter passthrough allows one to use custom variables and/or functions as parameters. For example we can use parameters of other gates:

    from mrmustard.lab.gates import Sgate, BSgate
    BS = BSgate(theta=np.pi/4, theta_trainable=True)[0,1]
    S0 = Sgate(r=BS.theta)[0]
    S1 = Sgate(r=-BS.theta)[1]
    circ = S0 >> S1 >> BS

    Another possibility is with functions:

    def my_r(x):
        return x**2
    x = math.new_variable(0.5, bounds = (None, None), name="x")
    def cost_fn():
      # note that my_r needs to be in the cost function
      # in order to track the gradient
      S = Sgate(r=my_r(x), theta_trainable=True)[0,1]
      return # some function of S
    opt.Optimize(cost_fn, by_optimizing=[x])


  • Adds the new trainable gate RealInterferometer: an interferometer that doesn’t mix the q and p quadratures. (#132)

  • Now marginals can be iterated over:

    for mode in state:


Breaking changes

  • The Parametrized and Training classes have been refactored: now trainable tensors are wrapped in an instance of the Parameter class. To define a set of parameters do

    from import Parametrized
    params = Parametrized(
        magnitude=10, magnitude_trainable=False, magnitude_bounds=None,
        angle=0.1, angle_trainable=True, angle_bounds=(-0.1,0.1)

    which will automatically define the properties magnitude and angle on the params object. To access the backend tensor defining the values of such parameters use the value property


    Gates will automatically be an instance of the Parametrized class, for example

    from mrmustard.lab import BSgate
    bs = BSgate(theta = 0.3, phi = 0.0, theta_trainable: True)
    # access params

    (#133), patch (#144) and (#158).


  • The Parametrized and Training classes have been refactored. The new training module has been added and with it the new Parameter class: now trainable tensors are being wrapped in an instance of Parameter. (#133), patch (#144)

  • The string representations of the Circuit and Transformation objects have been improved: the Circuit.__repr__ method now produces a string that can be used to generate a circuit in an identical state (same gates and parameters), the Transformation.__str__ and objects inheriting from it now prints the name, memory location of the object as well as the modes of the circuit in which the transformation is acting on. The _markdown_repr_ has been implemented and on a jupyter notebook produces a table with valuable information of the Transformation objects. (#141)

  • Add the argument ‘modes’ to the Interferometer operation to indicate which modes the Interferometer is applied to. (#121)

Bug fixes

  • Fixed a bug in the State.ket() method. An attribute was called with a typo in its name. (#135)

  • The math.dagger function applying the hermitian conjugate to an operator was incorrectly transposing the indices of the input tensor. Now math.dagger appropriately calculates the Hermitian conjugate of an operator. (#156)

  • The application of a Choi operator to a density matrix was resulting in a transposed dm. Now the order of the indices in the application of a choi operator to dm and ket is correct. (#188)


  • The centralized Xanadu Sphinx Theme is now used to style the Sphinx documentation. (#126)

  • The documentation now contains the section. The optimization examples on the README and Basic API Reference section have been updated to use the latest API. (#133)


This release contains contributions from (in alphabetical order):

Mikhail Andrenkov, Sebastian Duque Mesa, Filippo Miatto, Yuan Yao

Release 0.2.0

New features since last release

  • Fidelity can now be calculated between two mixed states. (#115)

  • A configurable logger module is added. (#107)

    from mrmustard.logger import create_logger
    logger = create_logger(__name__)
    logger.warning("Warning message")


  • The tensorflow and torch backend adhere to MathInterface. (#103)

Bug fixes

  • Setting the modes on which detectors and state acts using modes kwarg or __getitem__ give consistent results. (#114)

  • Lists are used instead of generators for indices in fidelity calculations. (#117)

  • A raised KeyboardInterrupt while on a optimization loop now stops the execution of the program. #105



This release contains contributions from (in alphabetical order):

Sebastián Duque, Theodor Isacsson, Filippo Miatto

Release 0.1.1

New features since last release

  • physics.normalize and physics.norm are now available. (#97)

  • State now has a norm property. (#97)

  • Can now override autocutoff in State by setting the cutoffs argument. (#97)

Improvements since last release

  • Renamed amplification argument of Amplifier to gain. (#97)

  • Improved __repr__ for State. (#97)

  • Added numba section in about(). (#97)

Bug fixes

  • Renamed “pytorch” to “torch” in mrmustard.__init__() so that torch can be imported correctly. (#97)

  • Fixed typos in State.primal(), State.__rmul__(). (#97)

  • Fixed a multimode bug in PNRDetector.__init__(). (#97)

  • Fixed a bug in normalization of Fock. (#97)

  • Fixed a bug in (#97)


This release contains contributions from (in alphabetical order):

Sebastián Duque, Filippo Miatto

Release 0.1.0

New features since last release

  • This is the initial public release.


This release contains contributions from (in alphabetical order):

Sebastián Duque, Zhi Han, Theodor Isacsson, Josh Izaac, Filippo Miatto, Nicolas Quesada