v1.13.0

What's Changed

Breaking Changes

• ham_hubbard_hardcore fix description and sign convention of hopping strength t.
• heisenberg_from_edges fix sign convention of magnetic field terms.
• the quimb.tensor submodule structure has been refactored with tn1d, tn2d, tn3d, and tnag submodules for better organization. Imports from old locations will still work, but are deprecated. Public classes and functions such as MatrixProductState are directly accessible from the top level quimb.tensor module as before.

Enhancements:

Major updates to splitting/decomposing individual tensors/arrays:

• add array_split and array_svals as the primary array-level entry points for matrix decomposition, consolidating dispatch logic that was previously internal to tensor_core.
• add register_split_driver and register_svals_driver decorators for registering custom matrix decomposition methods with array_split and array_svals.
• allow array_split to handle batches of matrices (for most methods).
• array_split: automatically detect and forward valid kwargs to underlying decomposition methods.
• tensor_split and array_split: expand absorb options significantly beyond "left", "both", "right", None to include "lorthog", "rorthog", "lfactor", "rfactor", "lsqrt", "rsqrt and "s" for returning partial results (single factors or singular values only). Default changed from "both" to "auto", which uses each method's natural default.
• add method "svd:eig" with main implementation svd_via_eig for efficient SVD via hermitian eigen-decomposition, with shortcuts for all absorb modes. This can be faster (especially e.g. on GPU) than the standard SVD, but entails some loss of precision.
• tensor_split: rename method option "eig" to "svd:eig" to make it clearer that this is an SVD split via eigen-decomposition. "eig" remains as a deprecated alias for "svd:eig".
• add method "svd:rand" with main implementation svd_rand_truncated for randomized SVD with truncation, with shortcuts for all absorb modes. (This is a new and backend agnostic implementation as opposed to the existing 'rsvd' method).
• add method "qr:cholesky" qr_via_cholesky for efficient QR or LQ like decompositions via cholesky decomposition, with shortcuts for all absorb modes. This can be faster than the standard QR (especially on GPU) but entails some loss of precision.
• tensor_split and array_split: add "lsqrt" and "rsqrt" absorb options, update cholesky decomposition to cholesky_regularized with shift as exposed parameter.
• compute_oblique_projectors: allow method kwarg.
• QR decomposition: add stabilize kwarg for controlling QR stabilization behavior.
• decomposition methods: various compatibility improvements for JAX backend.

Other enhancements:

• add shift and clock operators.
• add Tensor.isfermionic and TensorNetwork.isfermionic methods.
• add Tensor.isblocksparse and TensorNetwork.isblocksparse methods.
• add phase_dual option to TensorNetwork.conj.
• rename tensor_network_1d_compress_zipup_first to tensor_network_1d_compress_zipup_oversample and standardise oversample arguments.
• add tensor_network_1d_compress_srcmps_oversample and tensor_network_1d_compress_fit_oversample methods.
• add connected_bipartitions for finding all connected bipartitions of a tensor network
• tn.distribute_exponent: add new_exponent option for specifying the new exponent value (default 0.0).
• tensor_network_1d_compress: correctly handle input networks with non-zero exponents and equalize_norms.
• add tensor_network_gate_sandwich_inds for applying a gate and its conjugate like $G A G^\dagger$ to a tensor network.
• tensor_network_ag_gate: add which="sandwich" option for applying a gate and its conjugate like $G A G^\dagger$ to a tensor network, default to this if the supplied tensor network is a TensorNetworkGenOperator.
• add function tensor_network_ag_gate_simple for applying a gate to an arbitrary geometry tensor network vector or operator, using simple update style gauges to perform any compression.
• insert_compressor_between_regions and upstream CTMRG/HOTRG methods: add explicit contract_opts, reduce_opts, and compress_opts keyword arguments for fine-grained control.
• TensorNetwork2D.contract_boundary, contract_ctmrg, [contract_hotrg](https://quimb.readthedocs.io/en/latest/autoapi/quimb/tensor/tn2d/core/index.html#quimb.tensor.tn2d.core.TensorNetwork2D....

v1.12.1

Breaking Changes

• bump minimum required python version to 3.11

Bug fixes:

• fix SimpleUpdateGen mixin inheritance order.
• fix insert_compressor_between_regions for fermionic tensor networks with bond signature +-.

Full Changelog: v1.12.0...v1.12.1

v1.12.0

What's Changed

Enhancements:

• move the experimental operatorbuilder module to the main quimb.operator module.
• add basic introduction to the operator module - "Basics"
• add new example on tracing tensor network functions ex_tracing_tn_functions
• tensor_split: add an info kwarg, supplying this with an empty dict or with the entry 'error' will store the truncation error when using method in {"svd", "eig"}.
• update infrastructure for TEBD and SimpleUpdate based algorithms.
• schematic.Drawing: add grid, grid3d, bezier, star, cross and zigzag methods.
• schematic.Drawing: add relative option to arrowhead, shorten option to text_between and text_left and text_right options to line.
• add Drawing.scale_figsize for automatically setting the absolute figsize based on placed elements.
• refactor TEBDGen and SimpleUpdateGen
• update the 2d specific SimpleUpdate to use the new infrastructure.
• tn.draw(): show abelian signature if using symmray arrays.
• tn.draw(): add adjust_lims option
• TNOptimizer: allow autodiff_backend="torch" with jit_fn=True to work with array backends with general pytree parameters, e.g. symmray arrays.
• tn.gen_gloops and tn.gen_gloops_sites: add join_overlap option. When building cluster by joining smaller generalized loops, this option controls how many nodes they need to overlap by to be joined together.
• all message passing routines: add callback option
• GBP: allow a message initilization function.
• D1BP: allow messages to be a callable initialization function.
• MatrixProductState.gate_nonlocal: add method="lazy" option for lazily applying a non-local gate as a sub-MPO without contraction or compression.
• LocalHamGen.apply_to_arrays: support pytree parameter arrays such as symmray.
• add Tensor.get_namespace and TensorNetwork.get_namespace for getting a reusable data array namespace
• TensorNetwork.isel: use take where possible to better support e.g. torch.vmap across amplitudes.
• MatrixProductState.measure, and MatrixProductState.sample: add backend_random option for specifying which backend to use for random number generation when sampling, this can be set for example to jax to make the whole process jittable, but by default is numpy, regardless of the actual array backend.
• add codes of 2-site MPS-MPO fitting for fermionic TN by @sjdu10 in #333

Bug fixes:

• fix insert_compressor_between_regions when insert_into is None.
• tensor network drawing, ensure hyper indices can be specified as output_inds.
• fix MatrixProductState.measure when using jax arrays ({issue}340).
• fix MatrixProductState.measure when projecting and keeping a site site ({issue}344).
• Relax equality tolerance for float32 values by @iyanmv in #329
• Use input parameters for optimize in tree als fitting by @h-larsson in #337

New Contributors

• @iyanmv made their first contribution in #329
• @sjdu10 made their first contribution in #333
• @h-larsson made their first contribution in #337

Full Changelog: v1.11.2...v1.12.0

v1.11.2

Enhancements:

• Update the introduction to tensor contraction docs
• Improve efficiency of 1D structured contractions when default optimize is used.

Bug fixes:

• fixes for MPS and MPO constructors when L=1, (#314)
• tensor splitting with absorb="left" now correctly marks left indices.
• tn.isel: fix bug when value could not be compared to string "r"
• truncated svd, make n_chi comparison more robust to different backends

Full Changelog: v1.11.1...v1.11.2

v1.11.1

Enhancements:

• add create_bond to tensor_canonize_bond and tensor_compress_bond for optionally creating a new bond between two tensors if they don't already share one. Add as a flag to TensorNetwork1DFlat.compress and related functions (#294).
• add ensure_bonds_exist for ensuring that all bonds in a 1D flat tensor network exist. Use this in the permute_arrays methods and optionally in the expand_bond_dimension method.
• tn.draw(): permit empty network, and allow color=True to automatically color all tags.
• tn.add_tag: add a record: Optional[dict] kwarg, to allow for easy rewinding of temporary tags without tracking the actual networks.
• add qu.plot as a quick wrapper for calling matplotlib.pyplot.plot with the quimb style.
• quimb.schematic: add zorder_delta kwarg for fine adjustments to layering of objects in approximately the same position.
• operatorbuilder: big performance improvements and fixes for building matrix representations including Z2 symmetry. Add default symmetry and sector options that can be overridden at build time. Add lazy (slow, matrix free) 'apply' method. Add pauli_decompose transformation. Add experimental PEPO builder for nearest neighbor operators. Add unit tests.

Bug fixes:

• Fix TensorNetwork2D.compute_plaquette_environments for mode="zipup" and other boundary contraction methods that use the generic 1D compression algorithms.
• parse_openqasm2_str allow custom gate names to start with the word gate (#312).
• MatrixProductState.gate_with_mpo: fix bug to do with inplace argument (#313).

Full Changelog: v1.11.0...v1.11.1

v1.11.0

Breaking Changes

• move belief propagation to quimb.tensor.belief_propagation
• calling tn.contract() when an non-zero value has been accrued into tn.exponent now automatically re-absorbs that exponent.
• binary tensor operations that would previously have errored now will align and broadcast

Enhancements:

• Tensor: make binary operations (+, -, *, /, **) automatically align and broadcast indices. This would previously error.
• MatrixProductState.measure: add a seed kwarg
• belief propagation, implement DIIS (direct inversion in the iterative subspace)
• belief propagation, unify various aspects such as message normalization and distance.
• belief propagation, add a plot method.
• belief propagation, add a contract_every option.
• HV1BP: vectorize both contraction and message initialization
• add qu.plot_multi_series_zoom for plotting multiple series with a zoomed inset, useful for various convergence plots such as BP
• add info option to tn.gauge_all_simple for tracking extra information such as number of iterations and max gauge diffs
• Tensor.gate: add transposed option
• TensorNetwork.contract: add strip_exponent option for return the mantissa and exponent (log10) separately. Compatible with contract_tags, contract_cumulative, contract_compressed sub modes.
• tensor_split: add matrix_svals option, if True any returned singular values are put into the diagonal of a matrix (by default, False, they are returned as a vector).
• add Tensor.new_ind_pair_diag for expanding an existing index into a pair of new indices, such that the diagonal of the new tensor on those indices is the old tensor.
• TNOptimizer: add 'cautious' ADAM
• TensorNetwork.pop_tensor: allow tid or tags to be specified.
• add an example notebook for converting hyper tensor networks to normal tensor networks, for approximate contraction - https://quimb.readthedocs.io/en/latest/examples/ex_htn_to_tn_2d.html#example-htn-to-2d
• Add SX and SXDG gates by @kevinsung in #277
• Add XXPLUSYY and XXMINUSYY gates by @kevinsung in #279
• Added progbar to qsim, and openqasm2 by @edenian in #288
• quimb.experimental.operatorbuilder: fix MPO building for congested operators (#296 and #301), allow arbitrary dtype (#289). Fix building of sparse and matrix representations for non-translationally symmetric operators and operators with trivial (all identity) terms.

Bug fixes:

• fix MatrixProductState.measure for cupy backend arrays (#276).
• fix linalg.expm dispatch (#275)
• fix 'dm' 1d compress method for disconnected subgraphs
• fix docs source lookup in quimb.tensor module
• fix raw gate copying in Circuit (#285)

New Contributors

• @edenian made their first contribution in #288

Full Changelog: v1.10.0...v1.11.0

v1.10.0

Enhancements

• tensor network fitting: add method="tree" for when ansatz is a tree - tensor_network_fit_tree
• tensor network fitting: fix method="als" for complex dtype networks
• tensor network fitting: allow method="als" to use a iterative solver suited to much larger tensors, by default a custom conjugate gradient implementation.
• tensor_network_distance and fitting: support hyper indices explicitly via output_inds kwarg
• add tn.make_overlap and tn.overlap for computing the overlap between two tensor networks, $\langle O |T \rangle$, with explicit handling of outer indices to address hyper networks. Add output_inds to tn.norm and tn.make_norm also, as well as the squared kwarg.
• replace all numba based paralellism (prange and parallel vectorize) with explicit thread pool based parallelism. Should be more reliable and no need to set NUMBA_NUM_THREADS anymore. Remove env var QUIMB_NUMBA_PAR.
• Circuit: add dtype and convert_eager options. dtype specifies what the computation should be performed in. convert_eager specifies whether to apply this (and any to_backend calls) as soon as gates are applied (the default for MPS circuit simulation) or just prior to contraction (the default for exact contraction simulation).
• tn.full_simplify: add check_zero (by default set of "auto") option which explicitly checks for zero tensor norms when equalizing norms to avoid log10(norm) resulting in -inf or nan. Since it creates a data dependency that breaks e.g. jax tracing, it is optional.
• schematic.Drawing: add shorten kwarg to line drawing and curve drawing and examples to the docs.
• TensorNetwork: add .backend and .dtype_name properties.

PRs:

• Circuit: add default dtype and convert_eager options by @jcmgray in #273
• add fit(method="tree") and fix ALS for complex TNs by @jcmgray in #274

Full Changelog: v1.9.0...v1.10.0

v1.9.0

Breaking Changes

• renamed MatrixProductState.partial_trace and MatrixProductState.ptr to MatrixProductState.partial_trace_to_mpo to avoid confusion with other partial_trace methods that usually produce a dense matrix.

Enhancements:

• add Circuit.sample_gate_by_gate and related methods CircuitMPS.reordered_gates_dfs_clustered and CircuitMPS.get_qubit_distances for sampling a circuit using the 'gate by gate' method introduced in https://arxiv.org/abs/2112.08499.
• add Circuit.draw for drawing a very simple circuit schematic.
• Circuit: by default turn on simplify_equalize_norms and use a group_size=10 for sampling. This should result in faster and more stable sampling.
• Circuit: use numpy.random.default_rng for random number generation.
• add qtn.circ_a2a_rand for generating random all-to-all circuits.
• expose qtn.edge_coloring as top level function and allow layers to be returned grouped.
• add docstring for tn.contract_compressed and by default pick up important settings from the supplied contraction path optimizer (max_bond and compress_late)
• add Tensor.rand_reduce for randomly removing a tensor index by contracting a random vector into it. One can also supply the value "r" to isel selectors to use this.
• add fit-zipup and fit-projector shorthand methods to the general 1d tensor network compression function
• add MatrixProductState.compute_local_expectation for computing many local expectations for a MPS at once, to match the interface for this method elsewhere. These can either be computed via canonicalization (method="canonical"), or via explicit left and right environment contraction (method="envs")
• specialize CircuitMPS.local_expectation to make use of the MPS form.
• add PEPS.product_state for constructing a PEPS representing a product state.
• add PEPS.vacuum for constructing a PEPS representing the vacuum state $|000\ldots0\rangle$.
• add PEPS.zeros for constructing a PEPS whose entries are all zero.
• tn.gauge_all_simple: improve scheduling and add damping and touched_tids options.
• qtn.SimpleUpdateGen: add gauge difference update checking and tol and equilibrate settings. Update .plot() method. Default to a small cutoff.
• add psi.sample_configuration_cluster for sampling a tensor network using the simple update or cluster style environment approximation.
• add the new example doc Exploring circuit sampling with various methods

v1.8.4

What's Changed

• fix MPS sample handling of RNG seed by @kevinsung in #248
• fix bug in applying MPO lazily to MPS (#246)

New Contributors

• @kevinsung made their first contribution in #248

Full Changelog: v1.8.3...v1.8.4

v1.8.3

Enhancements:

• support for numpy v2.0 and scipy v1.14
• add MPS sampling: MatrixProductState.sample_configuration and MatrixProductState.sample (generating multiple samples) and use these for CircuitMPS.sample and CircuitPermMPS.sample.
• add basic plot method for [SimpleUpdate] classes
• add edges_1d_chain for generating 1D chain edges
• operatorbuilder: better coefficient placement for long range / non-local MPO building

Full Changelog: v1.8.2...v1.8.3