Release Notes

Features

• Qobj.data_as now accepts "csr_array" and "dia_array", returning modern SciPy sparse arrays alongside the existing matrix formats. (#2938, by Veronika Kurth (@veronikakurth))

Performance

• Optimize "rouchon" method for stochastic solvers. (#2936, by Eric Giguère (@Ericgig))

Bug Fixes

• Use normalized mixed-state weights for trajectory allocation in mcsolve so that unnormalized density matrices correctly allocate the requested number of trajectories while preserving the original state trace. (#2927, by Rahul Shastri (@RahulShastri003))
• Cache _ishp, _iscp, _istp and _iscptp on Qobj, and invalidate all affected caches when dims, superrep or data are mutated. Fixes stale-cache bug where accessing type before reassigning dims/superrep would cause subsequent calls to return incorrect results. (#2886, by Alexis Le Gall (@QubitGoblin31))
• Fix to_choi and to_chi raising AttributeError instead of TypeError for unsupported Qobj types. (#2930, by Mike German (@steps-re))
• Fix a batch of small correctness bugs: Qobj.logm now reports "logm" in its non-square error, HEOMResult.final_ado_state returns the ADO state instead of the density matrix, qdiags no longer marks sub-unit-modulus diagonals as unitary, qzero/qzero_like no longer mark non-square operators as Hermitian, and the scipy lsoda excess-work message has a missing space restored. (#2931, by Mike German (@steps-re))
• Increase qfunc validity range to show coherent(alpha>=13) properly. (#2937, by Eric Giguère (@EricGig), reported by Dong-Sheng Liu (@DS-Liu) in #2932)
• Accessing HEOMResult.final_ado_state no longer raises an AttributeError when store_ados is not set; the attribute is now always initialized and returns None when ADOs were not stored. (#2933, by Mike German (@steps-re))
• Fix an integer overflow in qutip.piqs.energy_degeneracy and qutip.piqs.state_degeneracy. Degeneracies larger than int64 are now returned as float so arrays keep a numeric (float64) dtype instead of falling back to object (e.g. piqs.block_matrix(80, elements="degeneracy") no longer raises). (#2948, by Udit Jain (@uditjainstjis))
• Fixed clebsch silently returning wrong values (and later nan or raising) for large angular momenta, caused by negative exponents turning exact integer factorial products into underflowing floats. (#2954, by Udit Jain (@uditjainstjis))
• Restore unintentionally dropped support of tuples for specifying outer dimensions. List stays the preferred data structure for this, but tuples are allowed. (#2955, by Veronika Kurth (@veronikakurth))
• Fix typehint for tensor() when passing iterables like lists, tuples, etc. (#2949, by Mayank Goel (@Mayank447))

Documentation

• Document how tensor_contract indices map to flattened Qobj.dims entries. (#2920, by Jeonghoon Lee (@jeonghoon-ad))
• Minor typo correction in direct_sum_sparse's docstring explaining the list/tuple syntax for dimensions. (#2947, by Paul Menczel (@pmenczel))
• Add a temporary link to the user survey in the navigation sidebar. (#2917, by Veronika Kurth (@veronikakurth))

Internal

• Migrate QuTiP's data layer from scipy.sparse.*_matrix types to scipy.sparse.*_array types internally while keeping user-facing behavior (via Qobj.data_as and extract functions) backward compatible. Adds a forward-compatibility guard so QuTiP won't break when SciPy fully removes the legacy matrix classes. (#2938, by Veronika Kurth (@veronikakurth))
• Ensure output of block_diag in qutip/piqs/piqs.py and its tests stays of type scipy.sparse.spmatrix to prevent a premature switch to scipy.sparse.sparray output before the sparray migration is complete. (#2925, by Veronika Kurth (@veronikakurth))
• Tighten the absolute tolerance of sesolve by one order of magnitude to make Floquet solver tests with randomized inputs more robust. (#2913, by Veronika Kurth (@veronikakurth))

Build & CI

• Add dependabot to update GitHub actions. (#2906, by Mayank Goel (@Mayank447))

Contributing

• Refine the PR template and contribution guideline to request filling out the AI tools usage disclosure also for cases when no AI tools were used. (#2943, by Veronika Kurth (@veronikakurth))