Release notes

v1.0.3.post2 (unreleased)

Relative to v1.0.3.post1, this release adds QFD, an eigenvalue method in a basis of time-evolved states, and new measurement, shot-allocation and basis-growth options for GCiM. It changes the default measurement of sampled ADAPT, the default pair preparation and quadratures of FixedGCIM and ADAPT, the default memory limit of the local simulator, the shared Run and Result archive formats and the saving of fermionic operators. It cannot load Run or Result archives written by v1.0.3.post1.

QFD

QFD (nwqlib.algorithms.QFD) estimates the smallest eigenvalue of a Hermitian operator \(A\) in the basis \(V^{p}|\psi\rangle\) of integer powers \(p\) of one fixed product-formula step \(V\) applied to a reference state \(|\psi\rangle\). It solves the Hamiltonian pencil of Parrish and McMahon, arXiv:1909.08925v1, Eqs. (5)–(8), with the measurement and projected solve of FixedGCIM, by exact readout or with grouped shots. You replace \(A\), the reference, the time step and the powers. By default \(V\) is built from the Pauli terms of \(A\), and evolution="fermionic_blocks" builds it from complete number-conserving fermionic blocks of a separate generator. chemistry_evolution_terms (in nwqlib.algorithms.qfd) builds such a generator from the integrals of build_gcim_chemistry_problem data. The QFD guide describes the workflow, its limits and what the result does not establish.

GCiM measurement and its defaults

PencilAcquisition, passed as acquisition= to FixedGCIM, ADAPT and QFD, chooses how \(H\) and \(S\) are measured (How matrix elements are obtained). Its defaults change what an existing sampled script measures:

• Sampled ADAPT measures grouped joint states by default (matrix_measurement="grouped"), with one setting per qubit-wise commuting group and part, as FixedGCIM does. v1.0.3.post1 measured one Hadamard test per Pauli term and part. PencilAcquisition(matrix_measurement="termwise", pair_preparation="controlled") keeps that measurement as a comparison.
• quadratures="auto" measures only the real parts of the off-diagonal pairs when a proof from the stored inputs shows that \(A\) and every state the method can prepare are real, and records the imaginary parts as exact zeros. Otherwise it measures both parts, which v1.0.3.post1 always did. Real-only measurement does not show that the imaginary parts are zero on a noisy device.
• pair_preparation="prefix" runs the blocks that both preparations of a pair begin with once without control and controls only the rest, with the same joint state. v1.0.3.post1 controlled both complete preparations of every pair, which "controlled" still does.

For ADAPT, pair_preparation="vacuum_reference" and "vacuum_restore" control only the reference preparation, when planning can show from the Pauli terms of every pool generator that its exponential maps the all-zero state to itself up to a phase. FixedGCIM and QFD refuse both values in quantum execution unless the Plan measures nothing.

Shot allocation from a pilot stage

sampling=PencilSampling(total_shots=..., pilot_fraction=...) of FixedGCIM, ADAPT and QFD splits a shot budget into a uniform pilot stage and a main stage whose shots per setting the pilot pencil decides. Only the main-stage measurements form the final \(H\) and \(S\), the pilot shots count toward the total, and result.shot_allocation records the decision (Pilot and main shot allocation). For ADAPT the two stages measure the basis that growth ended with, with a budget separate from the shots= of the growth. Without sampling, every setting gets shots shots, as before.

ADAPT basis growth

ADAPT(basis_growth=BoundedRitzGrowth(max_chain_length=L)) grows a basis of generator chains of at most \(L\) generators, one chain per round, chosen by a score from the current Ritz pair, and basis_depth_limit bounds the CX count of each chain (Bounded Ritz growth). It changes the trial space, so its energies are those of a different projected problem from the default growth. With shots, an extension is scored only when it passes an empirical eligibility rule on the sampled overlaps, and growth stops with noise_inconclusive when none of a round's extensions passes. The default growth, basis_growth=None, is unchanged.

Circuit depth on IBM Runtime

max_native_two_qubit_depth=D of FixedGCIM, ADAPT and QFD refuses, before its submission, every circuit whose two-qubit depth after compilation for the device exceeds \(D\) or is not qualified (Native two-qubit depth). Only IBM Runtime computes this depth, and it qualifies it only for counts circuits with dynamical decoupling and twirling off, under qiskit-ibm-runtime 0.49.0. With any other installed version a depth limit refuses every IBM circuit, so after an upgrade of qiskit-ibm-runtime the limit is accepted again only once the re-check in Maintenance has qualified the new version. The default None sets no limit.

Other changes

• The default ExecutionLimits.simulator_memory_mb is 8192 MiB, raised from 1024. It is an upper limit that the memory needs of a local simulation are checked against.
• ProjectedPencil.offset_free_eigenvalues records the lowest one or two Ritz values before the identity coefficient is added back (Read the result).
• nwqlib.estimate(plan) of an ADAPT Plan, and of a FixedGCIM or QFD Plan with sampling, reports the counts that depend on the iteration or on the pilot as unavailable, with the reason that the new Method.query_catalog_reason() returns (ADAPT queries).
• The shared archive now saves an ordered fermionic operator, such as one from ingest_fermion, in its own arrays. v1.0.3.post1 raised ValueError when an archive had to save one.
• The coefficient_cutoff documentation of build_gcim_chemistry_problem now states that OpenFermion's operator addition, which its Jordan–Wigner transform also uses, applies OpenFermion's own tolerance, in OpenFermion 1.8.1 deleting a term whose running coefficient has absolute value below 1e-8. A stored Pauli coefficient can therefore differ from the sum of its integral contributions by those deletions.
• The new example notebook gcim_qfd_shallow_workflows compares the measurement and basis options of GCiM and QFD on H2 and a stretched H4 chain and plans a 26-qubit model. The eigenvalue notebook adds QFD and the grouped and termwise finite-shot counts, and the resource notebook gives the real-only GCiM settings, shots and CX bound. The stored outputs of every example notebook come from this version, including the 8192 MiB memory default where a notebook prints it.

Formats

The shared archive formats are nwqlib.run/22 and nwqlib.result/15, and this release cannot load nwqlib.run/21 or nwqlib.result/14 archives written by v1.0.3.post1. Keep v1.0.3.post1 available to read those archives or finish their pending runs, or replan and rerun with this release. The Plan and Result content hashes of existing FixedGCIM and ADAPT configurations change, because their records gain the new fields with their defaults, and no reader for the earlier hashes is kept.

The method archives embedded in these files change from fixed_gcim/6 to fixed_gcim/7 for FixedGCIM and from adapt/7 to adapt/8 for ADAPT. The new QFD method archive uses qfd/1.

Package and citation source versions are 1.0.3.post2.

v1.0.3.post1 (2026-10-21)

Relative to v1.0.3, this maintenance release changes the test suite, the documentation, two extension interfaces, the shared Run and Result archive formats, the error for a missing compiler extra, the CX and rotation counts that follow Qiskit's gate definitions, QHD's projector choice at 7 qubits and the minimum Qiskit and qiskit-aer versions. It cannot load Run or Result archives written by v1.0.3. Default option values are unchanged.

The test suite holds 1,501 tests, against 3,259 in v1.0.3. It keeps the checks of computed quantities against independent references, the resource limits that refuse work before it starts, the normal save, load and resume paths, the checks against installed SDKs, the example notebooks and every test that the mutation probes or the documentation name. Tests were removed when they were judged to check only message wording, printed text, internal records and helper calls, the rejection of misuse such as unknown options or edited saved files, or a case that another current test covers.

New tests follow a user's workflow through the public API:

• The Python examples on the README and documentation pages run as written. On 20 of these pages, the values the examples compute are compared with independent computations.
• The nwqlib command lists the registered Methods, reports the options each Method requires as its constructor does, and lists installed third-party Methods without importing them.
• LCHS and QLS solutions for a periodic stencil agree with the dense matrix exponential and linear solve within their stated error bounds. On Aer, a QLS solution scales with the magnitude, sign and phase of the right-hand side, and its physical scalars lie within their error bounds. A two-qubit Pauli operator A with a supplied kappa gives the same solution as its dense matrix, sampled indices follow the solution's distribution, and each accuracy setting in a comparison meets its own tolerance.
• A classical LCHS solution with a constant source lies within its refined error bound of the closed-form solution, and a sampled expectation value lies within its reported sampling radius.
• The Solution of a LinearSystem or LinearDynamics carries the Problem's unit, and its NormSquared carries the square of that unit. An Expectation's unit labels its normalized expectation, so a QuadraticForm output has an unspecified unit unless it declares one.
• QCELS estimates the H2 ground energy from a Pauli Hamiltonian with and without a constant offset. Lanczos keeps only the reachable directions when the requested Krylov dimension exceeds the subspace that the initial state generates. A saved sampled Lanczos result reanalyzes with its shot counts, and scan by requested shots and logical_width keeps both Plans of a shots-width trade-off in its nondominated set.

The LCHS guide states that A must be a dense matrix or a periodic stencil, and how to pass a SciPy sparse matrix or a Qiskit SparsePauliOp in dense form. The pages on comparing and ranking Plans state that ranking sampled Plans by width needs ResourceContext(batch_schedule="serial"). The Units section states that a unit-normalized StateVector is dimensionless for every Problem. The maintenance guide describes the documentation example check.

The documentation wording has been revised.

Two extension interfaces and the shared archive formats change:

• NWQLib passes settings to the Method's prepare(plan, *, run, settings) hook on every call, so an override of that hook must accept the keyword. In v1.0.3, NWQLib passed settings only for "all", and the inherited Method.prepare_all_refusal refused "all" for an override without the keyword. The inherited Method.prepare_all_refusal now returns None, so a Method that cannot prepare every setting in advance overrides it, as Adaptive Methods describes.
• The prepares_from_templates attribute of backend adapters and the template_seed fields of NativePreparation and PreparedArtifact are removed. No backend included with NWQLib set that attribute. An adapter that set it no longer receives the construction_id, backend_context and template_seed keyword arguments at prepare or prepare_local. These arguments have no replacement.
• Because preparation records change, the shared archive formats are nwqlib.run/21 and nwqlib.result/14. This release cannot load earlier shared formats, including those written by v1.0.3. Keep the writer's NWQLib version available to read old results or finish pending runs, or replan and rerun with this release to create new archives. This also applies to Result and Run subfolders of QHD refinement and augmented-Lagrangian archives.

Without the nwqec extra, compile_logical raises ModuleNotFoundError with the install command before it starts its child process, and so does estimate_physical without the qre extra. In v1.0.3, the child process failed instead, and the error arrived as a RuntimeError that contained the child's traceback.

Every extra that installs Qiskit now requires Qiskit 2.6.0 or later, and the aer extra also requires qiskit-aer 0.18.0 or later. Upgrade with the extras you use, for example python -m pip install -U "nwqlib[aer]", so that pip also upgrades Qiskit and qiskit-aer. With older Qiskit or qiskit-aer releases installed, the counts described below do not match the installed synthesis, and exact Aer readouts list "unchecked qiskit-aer version". The validated environment in docs/ENVIRONMENT_LOCK.txt changes only these two packages.

Without ancillas, Qiskit 2.6.0 synthesizes a multi-controlled X gate with 5 to 32 controls, and a multi-controlled phase gate with 5 or more controls, by the construction of da Silva and Park (Phys. Rev. A 106, 042602), with 4k**2 - 4k + 2 CX for k controls. Qiskit 2.5.2 used more CX for these gates, for example 2,328 instead of 1,522 for an X gate with 20 controls. NWQLib's resource counts of the circuits that Qiskit builds follow the new definitions:

• The CX table of multi-controlled X gates, MCX_CX_BY_CONTROLS, has the new counts for 5 to 32 controls. The QLS CX estimates of projector flips and of the Q_b' and A_t predicates read it, and so do the LCHS SELECT bounds.
• The LCHS CX bounds count the CX of controlled phases and controlled U gates by the new definitions. The projector flip of a QSP pass with k block ancillas, controlled on its parity qubit, costs 36k**2 - 34k + 20 CX from k = 5, for example 750 at k = 5 against 746 with Qiskit 2.5.2, and 35,796 against 65,768 at k = 32.
• QHD's structured number projector uses Qiskit's multi-controlled phase from 7 qubits, where it costs 122 CX against 126 for the phase diagonal, instead of from 8 qubits. Its CX and rotation counts follow the new construction. At 8 qubits it costs 170 CX against 220 with Qiskit 2.5.2, and for a generic angle it has 135 arbitrary rotations and 44 exact T gates, against 147 and 116.

Qiskit 2.6.0 also corrects the global phase that the two-qubit peephole optimization of optimization levels 2 and 3 changed by pi for some CX patterns. The example notebooks were executed again with Qiskit 2.6.0 and qiskit-aer 0.18.0.

The probability windows of exact Aer readouts name qiskit-aer 0.18.0 as the checked release. Its statevector, fusion and gate-matrix sources are those of 0.17.2, so the windows themselves are unchanged.

On Linux, qiskit-aer 0.18.0 and PySCF bundle different builds of the OpenMP runtime, and a Python process that loads both can fail when it loads the second, with "cannot allocate memory in static TLS block". On the Ubuntu 24.04 hosts of NWQLib's CI, such a process failed with qiskit-aer 0.18.0 and ran with qiskit-aer 0.17.2. A process that uses the aer and chemistry extras together on Linux, such as the eigenvalue notebook, should therefore start with the OpenMP runtime of PySCF preloaded, as Dependency issues describes.

Package and citation source versions are 1.0.3.post1, with the project concept DOI.

v1.0.3 (2026-10-19)

Relative to v1.0.2, this release includes the mathematical and planning updates first published in 1.0.2.post1 and the complete BSD-2-Clause license with the PNNL/DOE disclaimer.

License and distribution

The root LICENSE contains the Battelle copyright notice, the two BSD redistribution conditions, and the PNNL/DOE disclaimer. Source and wheel distributions carry this file unchanged. The publishing workflow compares the packaged license with the repository file byte for byte before uploading to PyPI. Zenodo metadata uses the explicit bsd-2-clause license identifier.

Mathematical and planning changes

• Periodic LCHS uses a Strang error coefficient derived for the order in which its circuit applies the operators. When steps are selected from a synthesis tolerance, pure diffusion on one or two system qubits selects one step with zero ideal product-formula synthesis error. Other error components, including quadrature and execution, remain separate.
• The relaxed_prefix Pauli bound uses weighted anticommutation degrees to tighten the previous suffix relaxation. The work and memory it counts against the limits include the additional degree passes and contraction work. A plan close to max_work or max_bytes can therefore be refused where v1.0.2 accepted it.
• The normalized-expectation error bound changes from 4*w*delta/r to the sharp 2*w*delta/r, where r is the norm of the computed physical vector, delta < r is its complete error budget, and w bounds the observable norm. Scaled evaluation avoids intermediate overflow and false zero from underflow. An unrepresentable positive bound is reported as unavailable evidence.
• Jacobi–Anger selection uses a tighter real-argument Bessel remainder and chooses the smallest usable polynomial degree within its finite search horizon. Ties select the smaller computed complete tail, then the earlier terminal. The search examines at most 202 terminals. A lower degree can have a larger reported tail. When the only passing candidates have an analytic remainder that underflows to zero, the polynomial remains available with unknown tail evidence. Phase fitting is unchanged.
• Five default planning limits increase as shown below. Independent byte and execution limits still apply, so passing the planning checks does not establish that a plan can be executed.

Planning limit	v1.0.2	v1.0.3
AdmissionLimits.max_definitions	16,384	1,000,000
FixedGCIM.max_admission_steps	1,000,000	10,000,000
LCHS.max_select_work	100,000,000	1,000,000,000
LCHS.max_dense_select_slots	256	4,096
LCHSRefinement.max_structural_work	100,000,000	1,000,000,000

Documentation and compatibility

The QHD documentation scopes the pre-expansion term-count bound to constants, symbols, addition, multiplication and nonnegative integer powers. General SymPy rewriting work and storage are outside that bound. The mathematics and algorithm guides clarify the distinction between component bounds and total error, and the LCHS introduction and resource-estimation notebooks have updated outputs.

Saved-format labels are unchanged from v1.0.2. Loading saved construction data preserves the choices recorded in it. Replanning can change step counts, polynomial degrees, Pauli-triangle bound variants (bound_variant) and content hashes. Reanalyzing saved LCHS observations uses the factor-2 normalized-expectation bound. Scalar error propagation and Bessel-tail evaluation use ordinary floating-point rounding.

v1.0.2.post1 (2026-10-08)

Relative to 1.0.2, this maintenance release tightens sufficient error bounds, improves scalar range handling and raises five planning limits. Saved format labels are unchanged. Replanning can select different step counts, polynomial degrees, Pauli-triangle bound variants (bound_variant) and content hashes. Loading saved construction data preserves the choices recorded in it.

• Periodic LCHS uses the matching-operator Strang coefficient for the factor order that its circuit applies. Pure diffusion on one or two system qubits has zero ideal product-formula synthesis error and selects one step. For larger grids its coefficient is no larger than the previous norm-cubed bound. It bounds the weighted finite-branch synthesis component. Kernel, quadrature, angle and execution components need their own error evidence.
• The pair-only relaxed_prefix Pauli bound uses weighted anticommutation degrees to limit nested mass. It encloses the full Pauli-triangle expression and is no larger than the previous suffix relaxation. The work and live storage it counts against the limits include the added degree passes and pair contraction. Because the relaxed count includes its degree passes, a configuration near max_work or max_bytes that 1.0.2 planned with relaxed_prefix can be refused. The full triangle expression can still exceed the actual evolution error.
• The normalized-expectation error bound uses the sharp factor 2*w*delta/r, for a complete supplied physical-vector budget delta < r and an observable norm bound w. Scale composition prevents avoidable intermediate overflow and false zero from underflow. A positive result outside the representable range has unavailable error evidence. This scalar propagation uses ordinary floating-point rounding, and unknown input components remain unknown. A new analysis of saved LCHS observations uses the factor-2 normalized-expectation bound.
• Jacobi–Anger selection uses a real-argument Bessel remainder and chooses the smallest usable degree within its existing finite horizon, then the smallest computed complete tail and the earliest terminal. The search examines at most 202 terminals. A lower degree can have a larger reported tail. Tail evaluation assumes ordinary floating-point rounding. If a selected polynomial's tail certificate is unrepresentable, the polynomial remains available with unknown tail evidence. Phase fitting and its convergence scope are unchanged.
• AdmissionLimits.max_definitions increases from 16,384 to 1,000,000, FixedGCIM.max_admission_steps from 1,000,000 to 10,000,000, LCHS.max_select_work from 100,000,000 to 1,000,000,000, LCHS.max_dense_select_slots from 256 to 4,096, and LCHSRefinement.max_structural_work from 100,000,000 to 1,000,000,000. Corresponding helper defaults agree. These limits allow more declarations, checking work, SELECT slots or counted construction work. Independent byte, native and execution limits still apply, and a planning count does not establish feasible execution.
• QHD's pre-expansion estimate is described as a nominal polynomial-skeleton charge. Its term-count proof covers constants, symbols, addition, multiplication and nonnegative integer powers. General SymPy rewriting, coefficient storage, symbolic temporaries and Sum/Product range work remain outside that proof. The symbolic decomposition and its scientific semantics are unchanged.
• Mathematical and algorithm documentation distinguishes quadrature-component looseness from total finite-action error, a hypothetical shot-based Gram cutoff from sampled data, and Low–Somma node count from coefficient mass and postselection probability.

Package and citation source versions are 1.0.2.post1, with the project concept DOI.

v1.0.2 (2026-10-07)

Relative to 1.0.1.post1, this release simplifies the library's internals, rewrites its documentation, and changes ten saved formats and the public names, arguments, fields and keys listed below.

Saved data

• The saved formats are qls/7, qpe/<estimator>/8, lchs/10, expectation/6, fixed_gcim/6, adapt/7, lanczos/7, qhd/8, nwqlib.run/20 and nwqlib.result/13. This release cannot load archives written under the earlier labels of these formats. The format labels of QHD refinement and augmented-Lagrangian directories are unchanged, but loading or resuming such a directory written by an earlier release is refused where it opens a nested Run or Result saved under an earlier label. A Program, evidence options, a QasmWriteReceipt, an IonQBackend configuration, a profile or a forecast saved as standalone JSON by an earlier release is refused by record validation. Keep the writer's NWQLib version available to read old results or finish pending runs, or replan and rerun with this release to create new archives.

Workflow and execution

• load_run no longer requires backend. When it is omitted or None, load_run reopens a classical Run and a Run on the default AerBackend(). A Run saved with any other backend still needs that backend, and a backend that does not match the saved configuration is refused with a message that names the saved configuration and what to pass.
• ObservationChunk has no readout(). Read chunk.observation.
• nwqlib.execution.prepare_experiment no longer takes reduction_allowance. An extension sets the work allowance of its reductions through its Method's reduction_allowance hook.
• python -m nwqlib.cli no longer runs the command line. Use python -m nwqlib or the nwqlib console script.
• check-method and check_method no longer write a wrong Result into a temporary saved archive to show that it is refused. They still run the in-memory wrong-answer check and the save-and-load round trip.
• third_party_registrations() no longer takes entries=. It discovers the installed entry points as before.
• ReducerOutput no longer takes index_width and has no payload_bytes, and nwqlib.core.planning.point_array_reservation is removed. These have no replacement. Registered reducers publish JSON values sized by dtype and shape.
• Measure has no basis field. Measurements are in the computational basis, and a basis change is a preceding block. A Plan whose Program contains a Measure therefore has a different content_id from the same Plan built by 1.0.1.post1.
• ProjectedVerificationOptions, EnergyShiftOptions and NumberSectorOptions no longer take source. Each class fixes its own source, so these options have different content_id values from those of 1.0.1.post1.

QASM export and import

• materialize_qasm3_file refuses a QasmWriteReceipt whose construction_id differs from the construction's content_id, before it opens the file. The message names both content hashes and asks for the QasmWriteReceipt that the writer returned for this construction.
• QasmMaterializationBudget no longer takes required_max_native_bytes or required_max_peak_rss, whose only accepted value was None. QasmMaterialization and QasmWriteReceipt no longer have native_bytes or peak_rss, which were always None. The importer reports no memory bound.

Eigenvalues and phases (QPE, Lanczos, GCiM)

• Classical QPE Plans no longer repeat the work and byte allowance of their classical routine as numerical_work and numerical_payload in plan.reconstruction. Read them from plan.construction.kernels[0].invocation_work and plan.construction.kernels[0].workspace[0].bytes. A quantum QPE Plan has no entry in plan.construction.kernels and recorded 0 in both fields.
• A supplied initial state that ADAPT, Lanczos or FixedGCIM pads to a power-of-two dimension now needs max_bytes of at least 64 bytes per padded entry, plus the supplied state's stored size, plus 65,536 bytes. Previously it needed 32 bytes per padded entry. At the default 10 GB limit, only supplied states with more than 2^26 entries are affected, and only ADAPT in classical execution pads a supplied state at that size.
• The refusal of a saved or reloaded FixedGCIM pencil whose overlap filter does not match its selected observations now reads "pencil overlap filter differs from its selected observations", in place of "pencil sampling policy differs from its selected observations".

Linear systems (QLS)

• QLS Plans no longer store encoding_id, preparation_id and encoding_operator in plan.reconstruction. A quantum Plan names its selected encoding and preparation in plan.construction.selections, and a classical Plan's encoding follows from plan.problem and plan.method.
• QLS work records (plan.reconstruction.work) no longer list matrix_products, which was always 0, in classical and quantum Plans alike. A classical QLS Result's applications[0].arguments no longer include it either.
• A Plan whose block encoding the Method selects, as quantum QLS does by default, now has the same content_id whenever it is built from the same inputs, in any process. An encoding that you wrap with select_block_encoding gets a new random label at each call, so Plans built from separate calls have different content_id values.

Linear dynamics (LCHS)

• nwqlib.algorithms.lchs no longer exports resolve_hamiltonian_evolution_backend or LCHS_HAMILTONIAN_EVOLUTION_IMPLEMENTATIONS. The LCHS.hamiltonian_evolution_backend field chooses the backend of each branch ("dense_exact", "trotter", "trotter_error_budgeted" or "qsp_block_encoding"), as its description states.
• LCHSAnalysis no longer has the references field, which was always "not_run". Reference checks are explicit verification calls.
• LCHSCoefficientPlan.record() no longer returns lchs_kernel_quadrature_compatibility or lchs_kernel_quadrature_certificate_status. Read the certificate status from solution_error_certificate_status. The compatibility key recorded only that an allowed provider combination was selected and has no replacement.
• LCHSQuadraturePlan and LCHSQuadratureData no longer have effective_range_k, which equaled range_k for every built-in quadrature rule, and lchs_quadrature_summary no longer returns the matching effective_K key. Read range_k, or the summary's K key.
• LCHSProductFormulaSelectPlan no longer has max_step_count. Read repetitions, which holds the same value.
• generate_lchs_product_formula_select_plan takes the quadrature data as quadrature= in place of matrix= and _quadrature_plan=. Build that data with generate_lchs_quadrature(matrix=..., final_time=..., method=...). The function still refuses a dimension that is not a power of two.
• nwqlib.algorithms.lchs.time_independent_common no longer exports complex_to_dict or is_power_of_two. They have no public replacement.
• The LCHS exact-readout work refusal gives the requested work, the limit and the minimum sufficient LCHS.max_readout_work, without the clauses about planning units and completed reduction units.

Optimization (QHD)

• QHD archives preserve unevaluated SymPy Sum expressions through Result loading and refinement or augmented-Lagrangian continuation.
• solve_augmented_lagrangian, plan_augmented_lagrangian, refine_box, constrained_grid_minimum and the ConstrainedQHDResult constructor no longer raise a dedicated TypeError for an argument of the wrong type. Such an argument now fails where the layer first uses it, typically with an AttributeError that names the missing attribute. refine_box still raises TypeError for a problem that is not an Optimization.
• The description of QHDVerification.max_bytes states that, for observed counts or probabilities without a kept state, the bytes that grid_minimum counts against the limit omit histogram construction or loading, decoding and passing-weight storage at every readout width. This limit therefore does not bound the minimum_success_mass step.

Subroutines

• The undocumented functions nwqlib.subroutines.lcu.lcu_preparation_implementation_metadata and nwqlib.subroutines.block_encoding.block_encoding_implementation_metadata are removed. Every block encoding records its selected implementation under BlockEncoding.metadata["implementation_metadata"], and every LCU circuit under LCUCircuit.preparation_metadata["implementation_metadata"]. The registries BLOCK_ENCODING_IMPLEMENTATIONS and LCU_PREPARATION_IMPLEMENTATIONS list each selectable implementation's record.
• coefficient_arithmetic is removed from TrotterStepSelection (its attribute and to_dict() key), from QPE reconstructions and from LCHS Trotter node records. NWQLib evaluates the coefficient one way, which the records' descriptions now state. bound_variant identifies the enclosed expression, and QPE reconstructions keep the exact coefficient in bound_coefficient.
• nwqlib.subroutines.trotterization.trotter_error_bound is removed. Call evaluate_trotter_bound(hamiltonian, time=t, steps=1, order=o), which returns the same bound under the same limits.
• Banded block encodings no longer repeat the band offsets as metadata["band_offsets"] or set metadata["toeplitz"], which was always True and has no replacement. Read the offsets from metadata["band_specification"]["offsets"].
• resolve_number_projector_lowering is removed from nwqlib.subroutines.hamiltonian_evolution.pauli_evolution. Call structured_number_projector_provider(support_size) from the same module, which returns the provider record that the removed function wrapped, with its CX count under "cx".
• nwqlib.subroutines.qsp.evolution no longer exports _reflection_about_ancilla_zero, the reflection I - 2|0><0| on the ancilla register. It has no public replacement. nwqlib.blocks.select_zero_reflection builds the reflection of the opposite sign, 2|0><0| - I.

Backends, device profiles and forecasts

• Resource forecasts compare by their declared fields, so saving and loading a PlanEstimate preserves equality regardless of cached lookups. PlanEstimate and AssessmentContext also allow the other object's equality comparison when its type differs.
• IonQBackend no longer takes debiasing. Requests are sent with debiasing disabled, as before.
• TimeModel and TimePrediction no longer have form. The model is the linear model that its coefficients, scope, domain and uncertainty describe.

Documentation and examples

• The guides, the API reference and the docstrings are rewritten to describe the behavior of this release, and the API reference renders the records it names.
• The resource_estimation_at_scale notebook is regenerated and executed with this release.

Package and citation versions are synchronized at 1.0.2.

v1.0.1.post1 (2026-10-04)

This release simplifies validation and saved-data loading, and removes forwarding APIs and their supporting code.

• Removes saved-record hash verification, unused ADAPT compiler hashes, QASM file digests and separate QPE/QLS Program checks. QHD loading compares the restored expressions directly with their saved descriptions.
• QASM import applies construction work limits to unchanged output from the same writer call and checks the file's actual size before reading.
• Use plan.to_record() and run.artifacts.get(manifest) in place of ArchiveFiles.write_plan and Run.hydrate. read_report returns the saved metadata without the metadata_validation entry.
• nwqlib.backends.capability_set, require_backend_capabilities and BackendTarget.missing_capabilities are removed. Use a tuple of capabilities when constructing a target and frozenset(required) - frozenset(target.capabilities or ()) to find missing capabilities. The adapter still checks the instructions and readout it can execute.
• QasmWriteReceipt.digest and QasmPrefix.digest are removed. bytes_written measures size, not content identity. Remove the digest key from a QasmWriteReceipt saved separately as JSON before calling QasmWriteReceipt.model_validate.
• The shared archive formats are nwqlib.run/19 and nwqlib.result/12. This release cannot load earlier shared formats, including those written by v1.0.1. Keep the writer's NWQLib version available to read old results or finish pending runs, or replan and rerun with this release to create new archives. This also applies to Result and Run subfolders of QHD refinement and augmented-Lagrangian archives. Package and citation versions are synchronized at 1.0.1.post1.

v1.0.1 (2026-10-03)

This release simplifies tests and the validation code they support. It removes 107 collected test cases while keeping independent checks of numerical results, resource limits and normal save, load and resume workflows.

• QLS uses the shared Program validation. Preparation and Qiskit circuit construction check their own work limits before any circuit is built, without a separate planning-time simulation of those checks.
• Several entry points stop explicitly rejecting incorrect shots types, unknown LCHS provider parameters, unknown analysis settings and unsupported chemistry reference names. Use the argument names and types listed in each Method's guide.
• GCiM analysis, Pauli grouping, the Pauli commutation tables of QPE, Trotter bounds and periodic LCHS stop computing additional bounds solely to recommend exact retry limits. Their work and memory checks remain in place.
• Result and Run loaders restore the current saved data with fewer repeated consistency checks. Save-time scientific associations, current format and array-layout checks, explicit Method selection and path restrictions continue to apply. Metadata reports expose the Result's association with its observations as metadata_validation["result_observation_association"].
• The standalone analytical reference case, its test data and its CI checks are removed. Its independent Lanczos moments and projected pencil are checked through the library's own readout and reconstruction functions.
• QPE sampling tests compare directly with an independently known eigenvalue. Controlled-gate cost tests share circuit construction while checking the distinct gate costs. Other tests drop fixed strategy choices, duplicate cases, exact SDK-version assertions and diagnostics for obvious input misuse. Helpers and documentation used only by those tests are removed with them.

v1.0.0.post3 (2026-10-02)

Counts-based QHD box refinement and augmented-Lagrangian runs with refinement now include selected region mass lower bounds in print(result) and result.report(). The default 95% confidence covers the configured run, and the mass is conditional on valid decoding of the level's backend-sampled distribution. The report compares Hoeffding and one-sided Clopper–Pearson bounds after allocating half the failure budget to each method.

BoxRefinementResult.report() now defaults to failure_probability=0.05, and ConstrainedQHDResult.report() accepts the same parameter. Pass None to omit the statistical report. Each completed level records region_axis_counts and region_count. The outer archive formats are qhd.refinement/4 and qhd.constrained/4, and the formats of the outer record of a run directory (controller.json) are qhd.refinement_run/6 and qhd.constrained_run/5. Earlier outer formats cannot be loaded or resumed by these readers.

v1.0.0.post2 (2026-10-01)

This release changes documentation, one error message and repository settings. Code that runs with v1.0.0.post1 runs unchanged, with the same results.

• The documentation is reorganized by reader, and old page addresses still work.
• The API reference is rewritten, with a summary, parameters and examples for its entries.
• The descriptions in the configuration schemas that nwqlib options prints have changed.
• The example notebooks use the documentation's terms, and their outputs come from a new run.
• The import error of lcu_state_preparation="mps_circuit" gives the install command of scikit_tt, which is not on PyPI.
• The CI, Full CI, Mutation Probes and Nexus Offline Qualification workflows give their GITHUB_TOKEN read-only access to the repository contents.
• Two test assertions that failed on Python 3.14 are removed. Both checked deeply nested JSON in a hand-edited result.json.

v1.0.0.post1 (2026-09-30)

The README is rewritten for the PyPI page. Nothing else changed.

v1.0.0 (2026-09-30)

NWQLib applies quantum algorithms to scientific inputs and returns the computed quantities with their methods, resources and evidence. This is the first public release. The pages linked below hold the details.

Algorithms

• Linear systems: QLS with the QSVT inverse polynomial and the Dalzell kernel shortcut.
• Linear dynamics: LCHS with the An–Childs–Lin, Cauchy and Low–Somma kernels on product-formula and QSP block-encoding routes.
• Eigenvalues and phases: Chebyshev Lanczos, fixed and adaptive GCiM, and QPE with QCELS, RWPE, SPE and RFE.
• Expectation values of finite Pauli sums: Expectation.
• Optimization over a box: QHD in one-hot and binary encodings, with box refinement and an augmented-Lagrangian route for equality and inequality constraints.

Subroutines (API): block encodings, LCU, QSP and QSVT phase selection, coherent QPE, Hamiltonian evolution, Trotterization with error budgets and Pauli commutation tables, state preparation including matrix product states, fermionic pools, and Pauli decomposition.

End to end

• One workflow for every method: plan selects the construction, estimate and compare give its resource counts before any circuit is built, solve runs it, and prepare and submit run it as a Run that can be saved, reopened and continued (Plan, compare and solve, Run on a backend, Continue an interrupted run).
• Resource estimates come from the gate-count formulas of the selected construction, are checked against compiled circuits at small size, and extend to sizes no simulator holds (Estimate resources).
• Every limit on work, memory and stored data is checked before the work starts, and a refusal names the parameter and the amount (Supply inputs).
• Results carry their error evidence and references. Saved Results and Runs reload without planning or measuring again, a reopened Run can continue its unfinished work, and result.analyze reanalyzes the saved observations with new settings without new measurements (Save, load and reanalyze results, Check accuracy and verify a result).
• Backends: local Aer, NWQ-Sim with Slurm submission, IBM Runtime, IonQ, Quantinuum Nexus, and QASM export (Choose a backend). A command line lists the methods and prints their cards (Use the command line).

Documentation: a quickstart, the user guides, one guide per algorithm with its source map from code to paper equations, the mathematics page with the proofs and resource formulas, the backend guides, the API reference, and the development pages from the contribution policy to the code tour and the engineering constants.

Examples (overview): eight executed notebooks. Four introductions cover a linear system, linear dynamics, eigenvalues and optimization. Three paper-derived studies cover collision relaxation with QLS, heat flow with a boundary penalty with LCHS, and constrained optimization with QHD. One notebook plans five problems at 80 to 100 qubits and estimates their resources without simulating a circuit.