This release builds on our ongoing efforts to provide high-performance solvers for the stochastic Schrödinger and stochastic master equations. It also introduces new methods for solving the Lindblad master equation, including higher-order Rouchon methods with adaptive time-stepping, and jump and diffusive Monte Carlo methods (which perform especially well on GPUs!). An important performance issue that we thought had been resolved in the latest release v0.3.2 has also been fixed (multiplication of a sparse qarray by a dense qarray on the left).

Congratulations and welcome to @gitgan78, @derekeverett and @AurelienGauffre, who have made their first contributions to the library! 👏

🔨 Breaking changes

• Renamed dq.operator_to_vector() to dq.vectorize() and dq.vector_to_operator() to dq.unvectorize().

🚀 Features

• New Rouchon2 and Rouchon3 methods for dq.mesolve() with support for adaptive time-stepping! (#967, #968, #986 and #991) These solvers are significantly faster than existing ones when a stiff term is present in the Hamiltonian (e.g. $a^{\dagger 3}a^3$) or in a jump operator (e.g. $a^2$). For instance, for a stiff $H=a^{\dagger 3}a^3$, x15 faster than Dopri5 for n=32 and x130 faster for n=64! 🏎️ (fixed rtol/atol precision per step, on a CPU, dense layout)

  

• New JumpMonteCarlo and DiffusiveMonteCarlo methods for dq.mesolve() (#952 and #992).

• Stochastic solvers

  • Improved the Event method for dq.jssesolve(): added the smart_sampling option to sample the no-click trajectory only once (#943 and #956) and the dtmax parameter to control click times precision (#990).
  • Added a new EulerJump method for dq.jssesolve() and dq.jsmesolve() (#956 and #969).
  • Added the Rouchon1 method for dq.dssesolve().

• Added forward-mode automatic differentiation support by passing gradient=dq.gradient.ForwardAutograd() to solvers, see ForwardAutograd (#945).

• Added support for time-dependent Liouvillian in dq.mepropagator() (#953).

• Added clip() (#935) and prefactor() methods to timeqarray (#974). This is the start of a work to ease the simulation of pulse sequences

  f = lambda t: jnp.cos(2.0 * jnp.pi * t)
  H = dq.modulated(f, dq.sigmax())
  H_clipped = H.clip(0, 1)

  

• Utilities

  • Added dq.plot.wigner_data() to plot pre-computed wigner functions (#966).
  • Added dq.xyz() utility, which returns the Pauli $\sigma_x$, $\sigma_y$ and $\sigma_z$ operators (#972).
  • Added dq.thermal_dm() utility to get the density matrix of a thermal state (#744).
  • Added dq.ground_dm() and dq.excited_dm() (#977).
  • Added an offsets optional argument to dq.asqarray() to make it JIT-compatible (#954).

🐛 Bugs

• Fixed matrix multiplication of sparse dia qarray by dense qarray on the left (#973).
• Fixed a bug for modulated time-dependent qarrays preventing specifying functions returning a scalar output (#955).
• Make method a non-static argument for all solvers (#951).
• Fixed dq.plot.grid() for a single subplot (#965).
• Fixed the default layout of dq.eye_like() and dq.zeros_like(), which was dense and is now the same as the argument (#979).
• Added support for Hermitian matrices in dq.norm() and fixed the documentation (#957).

📖 Documentation

• New advanced tutorial on Continuous jump measurement.

  

📦 Other changes

• Lifted the deprecated upper JAX version constraint, which was set to <0.5.3 in the previous release v0.3.2 (#1000).

Full Changelog: v0.3.2...v0.3.3