Skip to content

Changelog

Included directly from the repository README.md so it never goes stale relative to the single source of truth.

v8.1.43

  • Added: seven standard-convenience functions Qiskit/PennyLane ship as everyday API that this package didn't: entangling_layer(n, pattern='linear'|'circular'|'full'|'star'|'brick'), pauli_expectation/pauli_sum_expectation (Pauli-string expectation values via O(dim) bit manipulation, never building the 2^n_qubits Hamiltonian matrix), ghz_state(n), sample_counts(sv, n_shots) (Qiskit-style get_counts()), statevector_fidelity(a, b) (the pure-state counterpart to uhlmann_fidelity, which only handles density matrices), qft(n, inverse=False, do_swaps=True), random_circuit(n_qubits, n_gates, ...), and draw_circuit(circuit, n_qubits) (plain-text diagrams). The first five were confirmed as real gaps, not speculative additions: the same patterns were found hand-duplicated with no shared, tested implementation across 20+ VQE/observable scripts and every test/experiment that preps a GHZ or Bell state, in this repo and the sibling Dense-Evolution-Ising-Tests. qft/random_circuit/draw_circuit round out the pass for completeness (no internal duplication motivated them the same way).
  • Real bug found and fixed while verifying pauli_expectation: the first implementation assumed qubit q is bit q of the basis-state index. DenseSVSimulator actually stores qubit 0 as the most significant bit (confirmed empirically and matches conftest.py's own patched measure(): phys_q = self.n - 1 - qubit_idx). Fixed and re-verified — 500 random 4-qubit states cross-checked against brute-force dense Pauli matrices (numpy.kron), exact match (max error 0.00e+00). qft was verified the same way: max error 1.4e-15 against the analytic DFT matrix across 1-4 qubits, plus an exact QFT-then-inverse-QFT round trip.
  • Real bug caught before commit: draw_circuit's first version used Unicode box-drawing characters (─│●), which crashed print() on a default Windows console (cp1252 can't encode them). Switched to plain ASCII (-|*) — a diagram meant for a terminal or log file has to survive whatever console encoding the caller has.
  • 72 new tests across test_topology.py/test_observables.py/test_states.py/test_measurement.py/test_qft.py/test_random_circuit.py/test_drawing.py.
  • Added: macos-latest to the CI test matrix (previously ubuntu-latest only), prompted by a direct question about macOS support that had never actually been verified. This surfaced two real, unrelated macOS-only process crashes, both fixed:
  • Qiskit itself segfaults on macOS CI runners. qiskit.circuit.QuantumCircuit.__init__ — the simplest possible call, QuantumCircuit(3) — reproducibly crashed the whole process (SIGSEGV) on Python 3.10/3.11/3.12, macos-latest (arm64), deterministically at the same line every time. Not a Dense-Evolution bug: every Dense-Evolution-only test passed cleanly on macOS before hitting this file. First fix attempt (skipif on the test class) didn't actually solve it — qiskit = pytest.importorskip('qiskit') is a class-body statement, and pytest runs class bodies at collection time regardless of a skip marker on the class, so qiskit was still being imported into the process either way, and the crash just moved to a non-deterministic segfault during interpreter shutdown instead. Real fix: guard the class definition itself behind if sys.platform == 'darwin', so the class body never executes there at all; a pytest.mark.skip-decorated empty stub stands in for visibility. TestPennyLaneInterop, same file, unaffected either way.
  • A second, unrelated shutdown-time segfault, independent of Qiskit. Confirmed by the above fix: with Qiskit fully removed from the macOS process, the exact same class of crash (SIGSEGV, exit 139) still happened on Python 3.10/macos-latest — every one of 523 tests passing, coverage.xml already written, pytest's own summary already printed, then a crash a few seconds later during Python interpreter finalization. Points to native-extension teardown (most likely JAX/XLA's runtime shutdown on macOS ARM, a known category of issue unrelated to this package's own code). Fixed with a trylast pytest_sessionfinish hook in conftest.py that calls os._exit(exitstatus) on macOS right after pytest's own work — and pytest-cov's coverage.xml write — are done, bypassing the interpreter-finalization phase that was crashing. No-op on Windows/Linux.
  • CI now runs 6 jobs (2 OSes x 3 Python versions) instead of 3; all green.

v8.1.42

  • Added: a full documentation site, tatopenn-cell.github.io/Dense-Evolution — MkDocs + Material, auto-deployed to GitHub Pages on every push to docs/**/mkdocs.yml/dense_evolution/** via .github/workflows/docs.yml. The full API reference (api/*.md, 11 module pages) is generated directly from this codebase's own docstrings via mkdocstrings (docstring_style: google), not hand-duplicated; the Changelog and License pages are single-sourced from this README and license.md via mkdocs-include-markdown-plugin so they can't drift out of sync. Two real bugs surfaced only by the strict-mode CI build, both fixed before the first deploy: a docstring bracket sequence in QASMCircuit.to_tuples misparsed as a broken markdown link by mkdocs-autorefs, and a Linux/Windows filesystem case-sensitivity mismatch (docs/LICENSE.md including ../LICENSE.md, but the tracked file is lowercase license.md) that only failed on the Linux CI runner, not locally.
  • First honest self-review pass (rereading the live site page by page rather than assuming it was finished) found and fixed 6 real gaps: the Getting Started Zero-Noise-Extrapolation example referenced undefined variables and couldn't actually be run as published; the dashboard command didn't mention app_dashboard.py only exists in a cloned checkout, not the pip package; there was no examples/tutorial page beyond the short Quick Start (added docs/examples.md — density-matrix ZNE healing, MPS for low-entanglement circuits, differentiable VQE, each adapted from already-tested code — experiments/matrix_healing_zne.py, test_mps.py, test_autodiff.py — not written fresh); related API pages had no cross-links (mitigation.md ↔ healing.md/registry.md, mps.md ↔ chunk.md/simulator.md); there was no favicon/logo (hand-authored docs/assets/favicon.svg, a qubit-orbit glyph in the site's own cyan accent); and there was no architecture diagram (added a Mermaid module-dependency graph to docs/index.md, traced from the real import statements in every dense_evolution/*.py file, not an idealized layering).
  • Second review pass found 3 more: API reference function signatures rendered as cramped single-line text because mkdocstrings needs Black or Ruff installed to format them (added ruff to the docs extras — confirmed via a before/after strict rebuild that signatures now render properly indented); this README's own ## ▍ Benchmarks section (real throughput/PennyLane-comparison numbers) had no path onto the site at all (added docs/benchmarks.md, single-sourced from this file); and CONTRIBUTING.md existed in the repo but wasn't linked from the site nav (added docs/contributing.md — also fixed two of its relative links, to license.md/SECURITY.md, to absolute GitHub URLs so they resolve correctly both on GitHub and once included into the docs site, the same class of case-sensitivity fragility as the earlier LICENSE.md bug).
  • Added: test_docs_examples.py, run in CI (ci.yml) on every push — locates each Python code block actually published in getting-started.md/examples.md by its section heading and executes it directly (not a hand-maintained copy), asserting on real output (ZNE fidelity improves, GHZ probabilities are exactly 0.5/0.5, VQE converges well below the Hamiltonian's mid-spectrum). A future signature change anywhere these examples touch now fails CI immediately instead of silently going stale until someone rereads the site by hand. Also brought CONTRIBUTING.md's own "Running tests" command back in sync with what CI actually runs — it had already drifted, missing four test files, before this release.
  • This README now links the docs site prominently (a docs badge alongside the existing CI/PyPI badges, plus a callout line right under "What It Is").

v8.1.41

  • Added: test coverage tracking via pytest-cov ([tool.coverage.run]/[tool.coverage.report] in pyproject.toml, a dev extras group, a coverage step in CI uploading coverage.xml as an artifact -- codecov badge/upload would need a CODECOV_TOKEN the user sets up themselves, not done here). Real measured project-wide coverage, honestly characterized before writing anything: an initial pass found 84.1%, driven up to 94.4% (449/450 tests passing, one isolated failure confirmed to be the same RAM-pressure artifact already documented for TestChunkMultiPiece, not a regression) by closing genuinely untested code paths across dashboard_core/md_telemetry.py (39.8%→100%), dashboard_core/qasm_library.py (61.1%→100%), dashboard_core/simulation_runner.py (64.4%→87.6%), dense_evolution/chunk.py (63.6%→95.9%), dense_evolution/registry.py (84.9%→99.5%), and dense_evolution/parser.py (87.6%→97.4%). dense_evolution/simulator.py (72.5%→78.6%) is limited by a confirmed coverage-tooling artifact, not a real gap: DenseSVSimulator.measure() is already thoroughly tested (TestMeasurement, 5 pre-existing tests) but coverage/pytest-cov fails to trace it correctly inside this specific large test suite -- confirmed directly with an isolated coverage.Coverage() script (bypassing pytest entirely) that the method traces correctly on its own, so no duplicate tests were added for it. dashboard_core/interactive_panel.py (82.0%) was deliberately left as-is, out of scope -- it overlaps with the separate, larger, not-yet-started "rebuild launch_interactive_panel with full Streamlit parity" task.
  • Real bug found and fixed along the way: dashboard_core.simulation_runner.estrai_valore_puro silently returned 0 for any numeric string input (e.g. estrai_valore_puro("3.5") returned 0, not 3.5) -- every Python str has a .index() method, so hasattr(elemento, 'index') was true for any string, routing it into the "object with an .index attribute" branch before ever reaching the isinstance(str) numeric-parsing block below it; calling .index() with no arguments always raised TypeError, silently caught and falling through to 0. If the QASM parser ever handed this function a string-typed qubit index or parameter, it was silently read as 0 instead of the real value. Fixed by moving the isinstance(str) check to run first.
  • chunk.py's biggest apparent gap (~200 lines, the distributed multi-device kernel) turned out to be a measurement gap, not a test gap: it's already covered by TestChunkDistributed, which only runs under XLA_FLAGS=--xla_force_host_platform_device_count=8 in a separate pytest invocation (same as CI already does) -- combining both invocations' coverage data gave the real 95.9% number.
  • 61 new tests total across test_dense_evolution.py and test_dashboard_core.py.

v8.1.40

  • Added: full jax.jit-compatible coverage for every function in dense_evolution.mitigation -- richardson_extrapolate_jit, zero_noise_extrapolation_jit (the predictive-healing branch), polynomial_extrapolate_jit, uhlmann_fidelity_jit, zne_density_matrix_jit. Each is a jit-safe _core (no np.iscomplexobj/np.asarray/float() calls on possibly-traced values -- those break jax.jit tracing) plus an unchanged eager wrapper, the same split already used by dense_evolution.mps's _jsd_vectors_jax/_jsd_vectors. Verified: each _jit variant matches its eager counterpart exactly; all compose correctly together inside a single outer jax.jit (the realistic use case -- several of these called in sequence inside a step function passed to jax.lax.scan, not each jitted in isolation). Real measured speedup on the full zne_density_matrix pipeline (not a single function): 4.5x-150x across 2x2-32x32 matrices, positive in every case tested.
  • Checked and rejected a suggestion (from the same external-AI source as v8.1.39's rewrite) to mark target_sigma_ideal static via functools.partial(static_argnames=...): calculate_delta_preemp already uses jnp.where internally, not a Python if, so it's trace-safe as a dynamic value -- marking it static would force a fresh XLA recompilation every time a caller varies it, with no benefit. Confirmed directly by testing both ways before deciding, not by assumption.
  • This closes out the density-matrix ZNE healing work for now. A further step -- wiring these functions into MPSSimulator so error mitigation can run on tensor-network-simulated circuits -- was scoped and intentionally deferred: it requires noise-channel support for MPS (doesn't exist yet, a large feature on its own -- MPO-based or trajectory-based), reduced (not global) density-matrix extraction via partial trace (also doesn't exist yet), and handling the multiple-noise-scale requirement ZNE needs, none of which reduce to "just call the existing functions." A global density matrix is the wrong target for large-qubit MPS use in the first place -- materializing a full 2^n x 2^n matrix defeats MPS's own reason for existing. Tracked separately, not started here.

v8.1.39

  • Changed: project_to_physical/uhlmann_fidelity rewritten as fully vectorized, jax.jit-compatible JAX (no Python while/for loop over eigenvalues) -- the previous versions had a dynamic Python loop that isn't traceable, forcing a host round-trip on every call. project_to_physical now uses Euclidean projection onto the probability simplex (Held, Wolfe & Crowder 1974; also Duchi et al. 2008) applied to the eigenvalues -- a different, fully array-vectorized algorithm for the exact same convex projection problem Smolin-Gambetta-Smith's paper solves (unique global minimum, so any correct algorithm must agree). uhlmann_fidelity now computes Tr(sqrt(inner)) as the sum of inner's eigenvalues' square roots instead of reconstructing the full matrix square root. Both verified numerically identical to the previous versions (~1e-15 max difference on the SGS paper's own worked example and 30 random test matrices) and confirmed to actually compile and run under jax.jit. No behavior change for any existing caller -- same inputs, same outputs, just usable inside a jitted pipeline (e.g. jax.lax.scan) going forward.
  • Originated as a suggestion from an external AI (Gemini), independently verified numerically before adopting -- same source also proposed reintroducing the predictive-healing coefficient perturbation (calculate_delta_preemp) for the density-matrix case as "the real improvement." That specific claim was rejected: it directly contradicts the already-measured result that the perturbation's effect is negligible (~0.14% coefficient shift even at a large observed coherence deviation, confirmed with a live numeric trace, not assumed) -- an external source getting a plausible-sounding architecture story right doesn't make an unverified empirical claim inside it right; each part was checked on its own.

v8.1.38

  • Documented: zne_density_matrix's docstring extended with a broader honest finding beyond v8.1.37's original single-configuration result. experiments/matrix_healing_zne_sweep.py (new) sweeps GHZ states from 2 to 5 qubits across all 5 NoiseModel channels (depolarizing, bitflip, phaseflip, amplitude_damping, combined). First pass (3 seeds, K=150 trajectories, 60 runs) looked mixed -- 54/60 positive, mean delta +0.09, but apparently unreliable for phaseflip/amplitude_damping (small or net-negative deltas at some qubit counts). Investigated further before trusting it: richardson_extrapolate's 3-point Lagrange coefficients (3, -3, 1) amplify statistical noise in their inputs (sum of squares 19x a single raw measurement's variance), so an undersampled density-matrix estimate makes the corrected result noisy even when the correction itself is sound -- confirmed directly by re-running the same "failing" configurations at higher K (300-1200 trajectories, more seeds), which turned them consistently and strongly positive. Re-ran the full sweep properly (K=400, 5 seeds, 100 runs total): 96/100 positive, mean delta +0.12, every single (qubit count, noise channel) combination net positive on average, growing to +0.20-0.25 at 5 qubits for depolarizing/bitflip. The 4 remaining negative runs are small (worst -0.02), consistent with residual Monte Carlo noise, not a systematic failure. No code behavior changed -- this is a documentation update, but the earlier "unreliable for phaseflip/amplitude_damping" wording in an intermediate draft of this entry was itself wrong and has been corrected here before ever reaching PyPI.
  • zne_density_matrix's docstring carries the corrected numbers plus a practical caveat for callers: correction quality depends on rho_at_scales being a low-noise (large-enough-K) estimate to begin with -- an inherent property of Richardson-style extrapolation's noise amplification, not a bug in this function.
  • Also tested: whether more noise-scale points (4, 5, or denser spacing) improve exact interpolation further. They don't -- more points make Lagrange coefficients larger (worse with denser spacing, a Runge's-phenomenon-like effect), amplifying noise further: mean delta drops from +0.148 (3 points, the original choice) to +0.081 (5 points, same spacing) to -0.220 (5 points, denser spacing -- actively worse than not correcting at all on some channels).
  • Added: dense_evolution.mitigation.polynomial_extrapolate -- least-squares polynomial extrapolation to zero noise, generalizing richardson_extrapolate. At exactly degree + 1 points it's mathematically identical to exact interpolation (verified to 1e-12); with MORE points it becomes an overdetermined fit that trades a little bias for real variance reduction instead of the exact interpolation instability above. zne_density_matrix now uses this (degree=2 default) instead of richardson_extrapolate -- at the standard 3-point setup this changes nothing (mathematically identical, all prior tests pass unchanged), but makes "pass more noise-scale points" safe instead of a trap.
  • Real, measured gain, fixed total measurement budget (experiments/matrix_healing_fixed_budget.py, the fair comparison -- splitting the same total number of Monte Carlo trajectories differently, not spending more): 3 points x K=400 (1200 total, classic 3-point ZNE) vs. 5 points x K=240 (1200 total, degree=2 fit) vs. 7 points x K=171 (~1200 total). 5 points matches or slightly beats the 3-point mean improvement (+0.150 vs +0.148) with 19% lower variance (std 0.050 vs 0.062) -- a free reliability gain at equal experimental cost. 7 points trades a little mean (+0.132) for 30% lower variance (std 0.043), a genuine tradeoff point.
  • Tested and rejected, for honesty: (1) applying zero_noise_extrapolation's existing predictive-healing coefficient perturbation (calculate_delta_preemp, originally designed for scalar coherence signals around target_sigma_ideal=10) to density matrices, using either matrix purity or Jensen-Shannon divergence (dense_evolution.mps._jsd_vectors) as the coherence signal -- no measurable effect even amplified 100x over its default strength. (2) Selecting the polynomial degree per-run via leave-one-out cross-validation on the noisy points themselves -- worse than a fixed degree (mean +0.113 vs +0.150, higher variance), because LOOCV on few noisy points is biased toward under-fitting. (3) Selecting degree via the same JSD coherence signal -- no exploitable correlation found between JSD and which degree actually performs best (degree 2 and 3 win about equally often across the whole measured JSD range). None of these are shipped; degree=2 fixed remains the best-tested default.

v8.1.37

  • Added: dense_evolution.mitigation.zne_density_matrix/project_to_physical/uhlmann_fidelity -- Zero-Noise Extrapolation extended from scalars/vectors to full density matrices, for noisy simulations tracked as ρ rather than a single statevector. project_to_physical implements Smolin, Gambetta & Smith's "Maximum Likelihood, Minimum Effort" (2012, arXiv:1106.5458) eigenvalue-projection algorithm (transcribed and checked against the paper's own worked numeric example before use) to correct the raw richardson_extrapolate output on a matrix stack -- polynomial extrapolation across noise-scaled density matrices is not itself guaranteed to be a valid (positive-semidefinite) density matrix, even when every input was. uhlmann_fidelity is a validation-only utility (F(ρ_A,ρ_B) = (Tr√(√ρ_A ρ_B √ρ_A))², reduces to |⟨ψ_A|ψ_B⟩|² for pure states, verified directly) -- deliberately not accepted as an input anywhere in the correction path, keeping "the ideal state is a grading criterion, not something the algorithm gets to see" structural rather than a convention callers have to remember on their own (the general principle: an error-mitigation technique that needs to know the answer already isn't one).
  • Honest, measured result (experiments/matrix_healing_zne.py, reproducible): 2-qubit Bell state, NoiseModel depolarizing noise at base_p=0.05, scales 1x/2x/3x, a 200-trajectory Monte Carlo density-matrix estimate per scale, averaged over 4 independent random seeds -- raw noisy fidelity ~0.865, corrected (extrapolated + projected) fidelity ~0.947, a real improvement of ~+0.08, positive individually on every seed tested (range +0.060 to +0.106). This is one measured data point in one noise regime, not a general guarantee -- other noise models, circuits, or noise strengths are untested and may behave differently; the docstrings say exactly this, not more.
  • This formalizes an experiment built directly on top of v8.1.36's richardson_extrapolate complex-dtype fix -- the same experiment run before that fix gave an invalid, misleadingly negative result (corrected fidelity worse than raw) purely because the bug was silently discarding the density matrices' imaginary parts, not because the technique doesn't work.

v8.1.36

  • Fixed: dense_evolution.mitigation.richardson_extrapolate (and zero_noise_extrapolation's healing-adapted branch) hardcoded dtype=jnp.float64 for expectation_values, silently discarding the imaginary part of complex input with only a low-signal ComplexWarning, no explicit error. Real inputs (the only case exercised by any existing caller) were unaffected, but any complex-valued caller -- e.g. extrapolating density-matrix entries, which are complex off-diagonal in general -- got a silently wrong, purely-real result. Found while building a density-matrix-healing experiment on top of this function; verified directly pre-fix (richardson_extrapolate([1+2j, 3+4j], ...) returned a purely real value, dropping real information). Fixed by picking values's dtype from the input itself (jnp.complex128 if np.iscomplexobj(expectation_values), else jnp.float64 -- identical behavior to before for every existing real-valued call site, verified: all prior tests pass unchanged).

v8.1.35

  • Added: MPSSimulator.run_circuit_jit(ops) -- a jax.lax.scan-fused, @jax.jit-compiled whole-circuit execution path for MPSSimulator, which previously had zero @jax.jit anywhere in the file: every gate ran as an eager JAX call, and the adaptive bond-dimension search inside _svd_truncate was a Python while loop forcing a host-device sync on every 2-qubit gate. Measured directly on a 60-qubit/428-gate stress circuit (the same one used to originally diagnose this): 88.9s (eager) -> 0.74s (fused, steady-state post-compile), ~120x faster, now within range of Qiskit Aer's MPS backend (0.64s) on the same circuit -- same physics exactly (chi_used=16, budget_violations=0, avg_JSD=0.0000, matching the eager run's own summary).
  • The two root causes, quantified before fixing: (1) zero JIT fusion -- confirmed via direct grep, not assumed; (2) non-adjacent 2-qubit gates expand into a SWAP chain at runtime, multiplying the real SVD-bearing gate count 2.4x on the stress circuit (187 logical 2-qubit gates -> 451 real apply_gate_2q calls, counted directly).
  • New technique (no precedent for this specific problem in chunk.py's own earlier JIT fusion, issue #5, whose geometry never changes gate-to-gate): the bond-dimension search is now a single vectorized pass (JSD computed for every candidate truncation size at once via jax.vmap, replacing the incrementing while loop) over gamma/lambda tensors padded to a fixed max_bond size always, with the real bond dimension tracked as a traced scalar and everything beyond it zero-masked rather than dynamically sliced (jax.lax.scan's carry requires constant shape/dtype every step). Non-adjacent gates are expanded into their SWAP-chain-equivalent adjacent-only op sequence at Python pre-compile time instead of via runtime dispatch, mirroring chunk.py::_compile_multi_chunk_ops's "all gate-identity branching happens before tracing starts" principle.
  • apply_gate_1q/apply_gate_2q/_apply_nonlocal_2q (the eager path) are unchanged and still available -- run_circuit_jit is a new, additional entry point, not a replacement. Explicit, intentional trade-off: the fused path pads every gamma/lambda to max_bond for the rest of the instance's lifetime after the call, trading this module's adaptive-memory benefit (its whole point for very large, low-entanglement circuits) for speed -- use the eager methods directly when memory, not speed, matters more.
  • Verified in 5 separately-checked stages, each against real data before moving to the next: (1) the vectorized bond-dimension search matches the eager while loop's (chi_new, jsd_val) exactly across 171 real cases from real circuits, including the budget-violation fallback; (2) the SWAP-chain pre-expansion matches the real runtime dispatch's exact call sequence across all 330 (q1, q2) pairs for 3 qubit counts, plus an end-to-end final-state replay check; (3) the fused kernel matches DenseSVSimulator on real entangling circuits (TVD < 1e-6) and matches the eager path bit-for-bit (fidelity 1.0 to machine precision) even under a genuine budget_violations truncation scenario, both dtypes; (4) entanglement_entropy/_bond_history/jsd_per_bond/truncation_errors/budget_violations end up populated identically to the eager path (machine-precision match), sourced from jax.lax.scan's stacked per-step diagnostics instead of Python list appends; (5) the real speed number above.
  • contract_to_statevector generalized (.squeeze(axis=0)/.squeeze(axis=-1) -> [0]-indexing) to work correctly on both the original unpadded boundary tensors (identical behavior, verified against the full existing test suite) and the new padded ones.

v8.1.34

  • Fixed: QASMParser._eval_param silently returned 0.0 for any gate-parameter expression it couldn't evaluate -- a malformed expression like rx(pi * / 2) q[0]; parsed successfully and silently produced rx(0.0), a different, valid circuit, with no signal a typo had happened. Found via an independent code-review report, reproduced directly (ast.parse('pi * / 2', mode='eval') itself correctly raises SyntaxError -- the bug was a blanket except Exception: return 0.0 around it, added when the previous raw-eval() code-execution vulnerability was fixed, that swallowed genuine syntax errors along with rejected/malicious expressions). Now raises ValueError instead, for both cases -- malformed expressions and disallowed/malicious ones alike (still structurally blocked by the same AST node-type whitelist either way, just explicit about it instead of silent). Same class of silent-wrong-behavior issue already fixed for unknown gate names and mismatched parameter batches (v8.1.27, issues #4/#6). 2 existing security tests (test_eval_param_blocks_sandbox_escapes, the full-parse exploit test) updated to expect the raised exception instead of a silent 0.0; 2 new regression tests added for the originally-reported malformed-expression case.

v8.1.33

  • Fixed: MPSSimulator._apply_nonlocal_2q silently swapped which qubit acted as gate-argument-1 vs gate-argument-2 whenever a non-adjacent 2-qubit gate (abs(q1-q2) != 1) was called with q1 > q2 -- e.g. apply_cx(ctrl=3, tgt=1) -- because the SWAP chain always ends up applying the gate at (min(q1,q2), min(q1,q2)+1) regardless of the caller's original argument order. For an asymmetric gate like CNOT this silently applied the wrong gate (control/target inverted) with no error or warning. Found via an independent code-review report, reproduced directly before trusting it: building a 4-qubit GHZ state via H(0), CX(0,3), CX(3,1), CX(1,2) gave fidelity 0.25 against the exact GHZ state ((|0000⟩+|1001⟩)/√2 instead of (|0000⟩+|1111⟩)/√2). Fixed with the same q1>q2 normalization apply_gate_2q's adjacent-qubit branch already used (transpose the gate's tensor axes, swap q1/q2) applied before the SWAP chain runs -- same fidelity check now reads 0.9999999657714559. Affects apply_cx/apply_cz/apply_swap/apply_ccx for any non-adjacent qubit pair called in decreasing order, reachable from dashboard_core's engine='mps' option on any QASM circuit with such a gate. 6 new regression tests in test_mps.py (direct repro, cross-check against DenseSVSimulator, non-adjacent unordered-control Toffoli, both float32/float64).
  • Checked, not changed: the same external report also flagged MPSSimulator.get_top_k_probable_states's amplitude summation as a potential bug -- independently re-verified against the exact contraction (matches to machine precision) and against the module's own documented history of already fixing exactly this class of bug. No live issue found. Its performance observation (the while loop in _svd_truncate isn't jax.jit-compiled) is real but is a speed question, not correctness, and is deliberately out of scope here.
  • Added: dashboard_core.run_md_telemetry now computes real molecular-dynamics-style telemetry (exact phase evolution under the selected diagonal Hamiltonian, plus dense_evolution's own NoiseModel driving a physically-grounded thermal-noise channel) whenever a compatible Hamiltonian and the circuit's current statevector are supplied, instead of always returning the synthetic placeholder data (run_md_simulation_dummy, documented as a placeholder since it was first written) that shipped every prior version. Falls back to the same labeled mock when no compatible Hamiltonian is active -- the result now carries df.attrs['is_real']/df.attrs['note'] so callers (and the UI) can show honestly whether a given MD panel is real or a demonstration, rather than presenting synthetic numbers as if they were physics. Energy/entropy/purity are all genuinely computed (purity via an ensemble-averaged density matrix, exact given the ensemble -- every LIBRERIA_HAMILTONIANE entry is ≤6 qubits, so this stays cheap). Wired into both ui_pages/quantum_simulator.py (Streamlit) and the new interactive panel below with a visible "DATI REALI" / "MOCK" badge on the MD tab.
  • Added: dashboard_core.launch_interactive_panel -- a full ipywidgets-based interactive dashboard for Colab/Jupyter, built directly on dashboard_core's existing functions (not a port of the old legacy/dash.py notebook export). Same panels as the Streamlit page (Overview, Fisica Stato, Mosaico, VQE Results, MD Results, Performance, 3D Helix, Hamiltonian, Mitigation (ZNE)) and the same sidebar controls (circuit source, engine, noise, ZNE/predictive healing, VQE, custom Hamiltonian, MD), driven by ipywidgets instead of Streamlit's rerun model -- stays inside a single notebook cell's output, no external tunnel/link. Requires the dashboard extra's ipywidgets>=8.0.0 (now included in that extra).

v8.1.32

  • Added: dashboard_core.mitigation_runner.run_mitigation_sweep/dashboard_core.mitigation_panel.build_panel_mitigation -- Zero-Noise Extrapolation and predictive healing, wired into the Streamlit dashboard as a real feature for the first time (dense_evolution.mitigation/dense_evolution.healing existed in the core package but had zero references anywhere in the dashboard before this). Runs the active circuit at 3 noise scales (1x/2x/3x the sidebar's noise probability) and extrapolates to zero noise via zero_noise_extrapolation, reused exactly as-is -- no reimplementation of the Richardson/healing math in the dashboard layer. New sidebar section ("🩹 Error Mitigation (ZNE)") and a new "Mitigation (ZNE)" tab showing fidelity-vs-noise-scale and ideal/raw/ZNE-corrected probability overlays; the healing-adapted path's coherence signal (sigma_at_base_noise) reuses the shot-noise binomial sigma already computed and shown elsewhere in the dashboard (build_panel_overview's NISQ Shot Histogram) as a pragmatic proxy, documented as such rather than presented as a first-principles derivation.
  • Fixed: dense_evolution.mitigation.richardson_extrapolate raised ValueError: Incompatible shapes for broadcasting whenever expectation_values[i] was itself array-valued (e.g. a full probability distribution per noise scale, not a bare scalar) -- found building the panel above, where each noise scale's "expectation value" is naturally a whole probability vector. Root cause: stacking expectation_values via jnp.asarray and multiplying by the (n,)-shaped Lagrange coefficients relied on JAX's default trailing-axis broadcast alignment, which pairs the coefficients against the last axis of the stacked array instead of the leading "one row per noise scale" axis. Fixed by reshaping the coefficients to broadcast against the leading axis explicitly and summing over axis=0 -- a no-op reshape for the pre-existing scalar case (verified: all prior scalar-input tests pass unchanged), now also correct for vector/array-valued inputs.
  • Fixed: dashboard_core.run_simulation's seed parameter only ever reached np.random.seed() (shot sampling) -- for a JAX statevector (the normal case), the actual noise channel applied via NoiseModel.apply_to_sv was never given rng=/jax_key=, so it silently fell back to OS-entropy randomness regardless of seed. Two calls with the identical seed produced different noisy results. Found as a flaky test (run_mitigation_sweep's 3-noise-scale sweep, same seed passed to each scale, expected -- and needed, for a meaningful extrapolation -- the same underlying noise realization scaled by p, got a fresh random one each time). Fixed by passing rng=np.random.default_rng(seed) explicitly to apply_to_sv.
  • Refactor: dashboard_core.py (1900-line single file) split into a package (dashboard_core/: qasm_library, hamiltonians, simulation_runner, vqe_engine, md_telemetry, plot_theme, metrics, panels, helix_3d, mitigation_runner, mitigation_panel, provenance) behind a backward-compatible __init__.py façade re-exporting the same public surface -- import dashboard_core as dc and every dc.xxx call site unchanged, verified via the existing test suite passing identically with zero test-file edits. _series_plot/_series_enhanced/_energy_enhanced (three near-duplicate time-series plotting functions, _energy_enhanced duplicating _series_enhanced's logic inline instead of calling it) unified into one parameterized _series_plot. Hardcoded color literals in build_panel_fisica/vqe_results/md_results/performance that exactly matched an existing plot_theme.C value now reference it instead of repeating the literal -- zero visual change, by construction; near-miss shades (close but not identical to a C[...] value) deliberately left alone rather than forced onto the "closest" key, which would have been a small but real, unauthorized color drift.
  • Also fixed: .github/workflows/ci.yml never ran test_mitigation.py or test_mps.py.
  • Refactor (ui_pages/, repo-only, not part of the pip package): the gradient header banner, the "run → session_state → guard → render" early-return pattern, and the neutral AI-shield metadata dict were each copy-pasted across 2-3 of quantum_simulator.py/vector_healing.py/quantum_scars.py/ai_middleware.py instead of shared. New ui_pages/components.py helpers (render_page_banner, render_run_guard, render_matplotlib_figure) and AI_SHIELD_NEUTRAL_META replace them -- same visuals, one place to change instead of three. Fixes, as a side effect, three independently-drifted hardcoded version numbers in the page banners (v8.1.9/v8.1.7/v8.1.22) plus a fourth in app_dashboard.py's own st.set_page_config/docstring -- all four now read dense_evolution.__version__ instead of a stale literal. Verified: existing test_ai_middleware.py/test_quantum_scars.py pass unchanged (22/22); all ui_pages modules import cleanly; the actual Streamlit server started and served all 3 page routes with no server-side exceptions in the log (no automated UI test exists for this layer, per the pre-existing documented gap -- this is the closest verification available without a browser screenshot tool in this environment).

v8.1.31

  • Changed: dashboard_core (the compute/panel layer behind the Streamlit dashboard) is now part of the installable package (pyproject.toml's packages), not repo-only -- pip install dense-evolution[dashboard,jax] alone now gives import dashboard_core as dc with the full 53-circuit QASM library and every panel builder, no git clone required. Superseded within the same release by v8.1.32's package split (dashboard_core.py → dashboard_core/) -- the packaging mechanism changed from a single py-modules entry to a real package directory, but the end result (pip install alone is sufficient) is unchanged.

v8.1.30

  • Added: Chunk.run_chunk_distributed -- dispatches the multi-chunk kernel across a real JAX device mesh (jax.shard_map + jax.lax.ppermute) instead of one process's RAM, one physical chunk per device (issue #1, v1 scope: jax.device_count() >= num_chunks). Cross-chunk gate mixing becomes point-to-point communication via ppermute, keyed on the fixed XOR-stride pairing between chunk indices -- the same pairwise-exchange pattern real distributed statevector simulators use; gates entirely local to a chunk, or with a chunk-select control and a local target (decidable from a device's own chunk index alone), need no communication at all. Verified for correctness against DenseSVSimulator on simulated multi-device CPU (XLA_FLAGS=--xla_force_host_platform_device_count=N) across all 6 gate/qubit-location cases individually, randomized mixed circuits at num_chunks in {4, 8}, and both dtypes -- CI now runs these in a dedicated step with 8 simulated devices (they skip cleanly on a single-device run). Real GPU-cluster network performance is not and cannot be measured in a single-device CI environment -- honest scope: this validates the sharding/communication logic is correct, not real multi-GPU throughput.
  • Fixed while building it: the ppermute permutation for each chunk-select qubit was built inside a Python list comprehension with the loop variable referenced directly in the lambda body -- classic late-binding closure bug, every branch silently used the last qubit's stride regardless of which was actually selected at runtime (only the branch for the last-iterated qubit happened to come out correct, which is exactly the failure pattern that surfaced it: q0/q1 wrong, q2 right, on a 3-chunk-select-qubit circuit). Fixed by binding the fully-precomputed static perm list as a default argument, evaluated eagerly at lambda-creation time instead of looked up by reference at call time.
  • Also fixed: .github/workflows/ci.yml never ran test_mitigation.py or test_mps.py -- both existed and passed locally but had never actually executed in CI. Added to the main test run.

v8.1.29

  • Added: dense_evolution.NoiseSpec -- a JAX PyTree (model/qubits static, p/jax_key as leaves, registered via jax.tree_util.register_pytree_node) that circuit_to_energy_fn's energy_fn now accepts as a fourth noise= argument, applying NoiseModel.apply_to_sv natively inside the same traced computation as theta (issue #8, scoped up from #7's narrower fix). Removes the need for an external, Python-side noise-application step and manual PRNG key bookkeeping around a training loop -- jax_key is a pytree leaf, so it flows through jax.jit/jax.grad/jax.vmap/jax.lax.scan the same way any other JAX array does, with no OS-entropy fallback (reproducibility is structural, not opt-in). Verified: reproducible from the same key, energy differs from the ideal circuit under nonzero p, gradient w.r.t. theta flows correctly through the noisy pipeline, jax.vmap over a batch of independent keys works with no external Python loop, jax.jit-wrapped result matches eager evaluation exactly.
  • Fixed: found while wiring the above -- NoiseModel.apply_to_sv's if model == 'ideal' or p <= 0.0: return sv early-exit raised TracerBoolConversionError as soon as p became a traced value (e.g. a NoiseSpec leaf under jax.jit) -- exactly the case #8 asked to enable. Fixed with a targeted try/except around only the p <= 0.0 optimization: every channel's math already reduces to a no-op at p=0 (fire = r < p is always False), so skipping the shortcut and falling through is still correct, it just loses the eager-mode fast path when p can't be checked concretely.
  • Note: noise.p is technically a differentiable pytree leaf, but the gradient through it is ~0 almost everywhere in practice -- the existing channels sample via a hard threshold (fire = r < p), not usefully differentiable without a smooth relaxation (e.g. Gumbel-softmax). Not addressed here, out of scope for what #8 asked (removing the state/key workaround, not making noise strength itself gradient-optimizable).

v8.1.28

  • Note: v8.1.27 was published to PyPI from a stale local checkout that predated the mitigation.py module and the #4/#6/#7 fixes below -- the installable v8.1.27 package on PyPI does not contain them, despite this changelog. PyPI doesn't allow re-uploading files under an already-published version, so this release exists specifically to ship the real content. If you're on 8.1.27, upgrade to 8.1.28 -- don't rely on 8.1.27's changelog matching what you actually have installed (same situation as v8.1.15, see that entry below).

v8.1.27

  • Added: dense_evolution.mitigation -- a Zero-Noise Extrapolation orchestrator (richardson_extrapolate, zero_noise_extrapolation, both exported from package root) using the field's standard ZNE vocabulary, composed on top of healing.py's existing primitives rather than duplicating or renaming them. Motivated by the ZNE+predictive-healing combination previously existing only as hand-written logic (_static_richardson/_adaptive_healing_richardson) inside a downstream repo's test file, calling calculate_delta_preemp as one ingredient with no reusable, standard-named entry point in the package itself -- undiscoverable by name for anyone (human or AI) reading dense_evolution looking for "ZNE". richardson_extrapolate generalizes to N arbitrary noise factors (verified: reduces exactly to the textbook (3, -3, 1) 3-point coefficients at noise_factors=(1,2,3), and is exact on synthetic linear data); the healing-adapted path is scoped to exactly 3 noise factors -- the only case it has been derived and verified against (cross-checked numerically against the original 3-point formula) -- and raises NotImplementedError rather than silently generalizing an unverified formula for other point counts.
  • Breaking: run_circuit/run_circuit_jit_beast_mode/run_parametric_batch_jit now raise ValueError on an unrecognized gate name instead of silently dropping the gate (issue #4). Verified pre-fix: run_circuit_jit_beast_mode([('h', 0), ('ch', 0, 1), ('x', 2)]) executed only h/x, ch vanished with no error or warning -- a typo in a gate name ('crx' instead of 'crz') used to silently run a different circuit than the one written. Anyone relying on the old silent-skip behavior will now get an explicit error instead -- there is no legitimate use case for a gate name silently doing nothing.
  • Breaking: run_parametric_batch_jit now validates parameter_batch.shape[1] against the actual number of parametric gates (rx/ry/rz/p/u1/phase/cp/crz/cphase) in base_circuit, raising ValueError on a mismatch instead of letting JAX's default out-of-bounds indexing clip silently (issue #6). This is the same positional-slot contract responsible for two independently-discovered corrupted-results bugs in a downstream repo (Dense-Evolution-Ising-Tests v2.1.0): a literal-float rotation gate still consumes a column, it is never exempt, and a grid built assuming otherwise used to produce a plausible-looking but wrong statevector with no signal that anything was off.
  • Fixed: NoiseModel.apply_to_sv used to silently ignore rng whenever sv was a JAX array, drawing from OS entropy instead -- a caller seeding rng = np.random.default_rng(42) for reproducibility got a different, non-reproducible result on every call, silently (issue #7). Verified pre-fix directly: two consecutive calls with the identical seeded rng and identical input diverged. Found while reviewing an end-to-end QML+noise+ZNE test script combining circuit_to_energy_fn, NoiseModel, and the new zero_noise_extrapolation under one jax.value_and_grad -- gradients did flow correctly through the whole pipeline (that part worked as intended), but the epoch-to-epoch loss instability in that script traced back to this, not to the optimizer or to mitigation.py. Fixed: when sv is a JAX array and jax_key isn't given explicitly, rng (if given) now derives the JAX key (rng.integers(...) seeds jax.random.PRNGKey) instead of being ignored -- a fresh, identically-seeded rng now reproduces the exact same sequence of keys across separate runs, matching the guarantee the NumPy path already gave. jax_key, when given explicitly, still takes precedence over a derived one. Making apply_to_sv itself composable inside a single jax.jit block (no internal Python-level branching or OS-entropy side effects) is a larger, separate redesign -- not done here, see issue #7 for the remaining scope.

v8.1.26

  • Verified, no bug found: two areas flagged as untested in a prior RAM-constrained environment -- Chunk's genuine multi-chunk dispatch (num_chunks > 1) on a 2-qubit gate spanning the very first (chunk-select) and very last (local) qubit of the register, and QASMParser's handling of for-loops with an unresolvable bound (an undeclared bound variable, so _resolve_int_expr returns None) -- both confirmed correct. Closed the real test-coverage gap: 4 new tests in TestChunkMultiPiece (the long-range 2-qubit gate case, plus MemoryPressureError actually firing on simulated low RAM for both the num_chunks==1 and num_chunks>1 code paths -- previously only ever exercised indirectly, never asserted on directly) and 6 new tests in TestQASMForLoop (unresolvable for bound stripped cleanly with following code preserved -- verified via probability comparison, not just the op list -- while blocks, multiple unresolvable constructs in sequence, and a resolvable for-unroll immediately followed by an unresolvable if-strip).

v8.1.25

  • Fixed: MPSSimulator._svd_truncate used to silently violate jsd_budget whenever max_bond was too small for the circuit's real entanglement -- the while jsd_val > self.jsd_budget and chi_new < max_possible loop exits with no signal once chi_new hits max_bond, even if jsd_val is still far above budget. summary()'s avg_JSD only reports the mean of the per-step local errors, not the accumulated global error, so it can read deceptively low while the final contracted state is badly wrong -- verified directly on an 8-qubit/15-layer entangling circuit with max_bond=2: TVD ~0.97 against DenseSVSimulator (a near-total mismatch) while avg_JSD read a reassuring-looking 0.0534. Added a budget_violations counter (incremented every time this happens, exposed in summary()) and a UserWarning on the first violation ("bond dimension capped at max_bond=..., jsd_budget=... not honored ... results may be unreliable") -- not an exception, so existing code that tolerates the tradeoff keeps working, but now with an explicit, checkable signal instead of a silently-optimistic average.

v8.1.24

  • Fixed: dense_evolution.healing.calculate_phi_ab raised ValueError: Clip received a complex value when called with complex statevectors (e.g. sim.get_statevector()) -- jnp.dot(semantic_change, ipg_vector) is the bilinear (non-conjugated) product on complex arrays and stays complex, which then hit jnp.clip at the end of the function. Fixed by using jnp.real(jnp.vdot(...)) -- the correct Hermitian-inner-product real part, identical to jnp.dot for the real-valued inputs every existing caller already uses (ia_utils.vector_healing.enhanced_dense_healing_hybrid), and now also correct for genuinely complex input. Verified against a manual NumPy Re(vdot(...)) calculation on a case with nonzero imaginary amplitudes (H+S gates), not just the crash repro (H+CX alone never produces a nonzero imaginary part, so it only proved "doesn't crash," not "computes the right number").
  • Fixed: ia_utils.vector_healing.median_healing/enhanced_dense_healing_hybrid emitted RuntimeWarning: Mean of empty slice whenever an input column was entirely NaN -- np.nanmean on an all-NaN slice returns NaN with a warning (silently caught and zeroed by the very next line, so the output was already correct, only the warning was noise). Fixed by pre-replacing whole all-NaN columns with 0.0 before calling nanmean, so it never sees an empty slice -- verified byte-identical output, zero warnings, in both functions (the preprocessing block was duplicated verbatim in each).
  • Fixed: enhanced_dense_healing_hybrid's fallback_triggered metadata flag used to reflect only the internal Phi-Trigger heuristic classifying a step as "static", regardless of whether the input actually contained any NaN/Inf -- verified directly that a clean, uncorrupted random Gaussian array (no corruption at all) still came back fallback_triggered=True. Root cause: the heuristic is tuned for trajectories with real underlying dynamics (verified separately against a real noisy VQE run, where it correctly stayed quiet at low noise and fired at high noise) and mistakes pure structureless IID noise -- which has no coherent trend to recognize as "genuine change" -- for anomalous static behavior on nearly every step. fallback_triggered is now gated on the original input actually containing NaN/Inf (checked before sanitization) in addition to the fallback having fired -- the dashboard's "Fallback scattato" indicator will now only light up for genuine NaN/Inf corruption, not general noise-driven Phi-Trigger corrections (those corrections still happen internally, they're just no longer mislabeled as a NaN/Inf fallback).
  • Docs: removed two README rows (kappa_stabilization, richardson_integration) describing functions that never existed in dense_evolution/healing.py (confirmed against the full 140-line file -- 7 functions plus MemoryReflectionEngine, neither name present anywhere), and corrected two code examples that still showed a median_fallback_threshold parameter on enhanced_dense_healing_hybrid that was removed from the actual function signature back when its corresponding UI slider was dropped (commit 69fc8a4) without updating the docs to match.

v8.1.23

  • Added: MPSSimulator (dense_evolution/mps.py, re-exported from the package root) -- a JAX-backed Matrix Product State simulator with adaptive SVD-truncated bond dimension. Verified exact against DenseSVSimulator on entangling circuits (TVD=0). Selectable as a dashboard engine (Quantum Simulator page) for circuits up to 24 qubits; for larger low-entanglement circuits, get_probabilities_sampled/get_top_k_probable_states scale to hundreds of qubits without ever materializing a (2**n,) statevector.
  • Fixed: large-qubit circuits submitted through the dashboard used to crash the whole Streamlit process (uncatchable OS-level OOM) instead of failing cleanly. They now route automatically through the existing Chunk anti-OOM wrapper, which raises a catchable, informative MemoryPressureError instead.
  • Changed (breaking): JAX is now a required core dependency (dependencies, not optional-dependencies) -- pip install dense-evolution installs it by default, no [jax] extra needed anymore. Every simulator backend in this package already required JAX in practice; the previous try/except ImportError numpy-fallback path was never a real, maintained alternative and is no longer reachable (the numpy code itself is left in place for reference, just dead). If you were pinning an environment without JAX and relying on degraded-numpy behavior, this release will break that -- install dense-evolution<8.1.23 to keep the old behavior.

v8.1.22

  • Added: ui_pages/quantum_scars.py — a new "Quantum Scars" dashboard page (app_dashboard.py's st.navigation), an interactive live demo of the PXP quantum many-body scar model (Rydberg blockade): builds H_PXP via sparse Pauli operators, exact-diagonalizes it (scipy.linalg.eigh, cached via st.cache_resource keyed on qubit count — the first use of that decorator in this codebase, since diagonalizing a dense 2**n_qubits matrix is genuinely expensive and depends only on that one slider), propagates a Néel initial state under real-time Hamiltonian evolution, injects real noise via NoiseModel.apply_to_sv, and lets you compare fidelity revival with no protection, a cheap constraint-subspace projection (no extra diagonalization needed — a combinatorial mask of which computational-basis bitstrings have no two adjacent 1-bits), or an idealized exact-eigenstate "tower" projection. Distills the investigation already published at quantum_scar_investigation — which found no genuine scar in Dense Evolution's own frustrated Ising grids (wrong observable + gauge equivalence), then validated the same verification pipeline against PXP, where scars are real and well documented — into something runnable instead of only readable. Verified via streamlit.testing.v1.AppTest (real button-click simulation, zero exceptions) and a new test_quantum_scars.py (17 tests, all passing) unit-testing the numerical core directly: validity-mask combinatorics, H_PXP Hermiticity and Hilbert dimension, fidelity=1/norm-preservation under propagation, and both projections staying normalized with zero weight outside their target subspace.

v8.1.21

  • Added: QASMCircuit.__iter__ — duck-types a parsed circuit as an iterable of the same tuples .to_tuples() returns, so it works anywhere a plain circuit list is expected (Chunk.run_chunk, QuantumTranspiler.transpile, ...) without remembering to call .to_tuples() first. Found via a user's own Colab testing: Chunk.run_chunk(QASMParser().parse(qasm)) — a very natural thing to try — raised TypeError: 'QASMCircuit' object is not iterable.
  • Fixed / Performance: Chunk's multi-chunk dispatch (num_chunks > 1) used to apply every gate through a Python loop calling non-JIT apply_gate_1q/apply_gate_2q — measured 6x slower than run_circuit_jit_beast_mode on an identical workload. Replaced with a single jax.lax.scan over the whole circuit operating directly on the stacked (num_chunks, chunk_dim) representation — never materializing a (2**n_qubits,) array, preserving the anti-OOM property Chunk exists for. The 6 gate/qubit-location cases were ported formula-for-formula from the old Python-loop implementation and verified case-by-case against it (all 18 pre-existing TestChunkMultiPiece tests, which cross-check against DenseSVSimulator, pass unchanged against the new kernel) before the old code was removed. Gate coverage is now built via GATE_IDS instead of the old GATES/PARAMETRIC_GATES lookup, aligning it with beast-mode's own coverage. Measured speedup on the exact benchmark that surfaced the problem (10 qubits/200 gates/4 forced chunks): 2.2s → 0.42s, now close to beast-mode's 0.37s on the same non-chunked workload.
  • Note: a Colab report of "17s vs milliseconds" that prompted this investigation turned out, on reproduction, to be num_chunks==1 (not the multi-chunk path at all) — 0.49s locally on the identical circuit, most likely first-time JIT compilation overhead on Colab's specific hardware rather than a code defect. The 6x slowdown that was real and is fixed here was found and confirmed with a separate synthetic benchmark (num_chunks forced via monkeypatch), not the original report.

v8.1.20

  • Fixed: from_pennylane/run_pennylane_circuit silently renumbered qubits whenever wires weren't touched in ascending order — both PennyLane's qml.to_openqasm and the older tape.to_openqasm() number exported QASM qubits by first-touch order, not by actual wire index (e.g. PauliX(wires=2) then CNOT(wires=[2,1]) exported as x q[0]; cx q[0],q[1];, silently mapping wire 2→q[0] and wire 1→q[1]). Found via independent fuzz testing (20 random circuits touching 4 wires in random order: 9/20 matched PennyLane's own results before the fix, 20/20 after). Fixed by passing an explicit wires= argument — the device's declared wire order for a QNode, the tape's own wires sorted ascending for a bare tape — forcing the true wire order into the export instead of relying on touch order.
  • Fixed: NoiseModel.apply_to_sv's depolarizing channel (and the combined channel's depolarizing sub-channel) picked which Pauli error (X/Y/Z) to apply using thresholds p/3 and 2p/3 compared against a draw uniform on the full [0,1) range — but that draw should only ever decide which Pauli fires, independent of the overall fire-rate p, so the thresholds needed to be the fixed values 1/3 and 2/3 instead. The bug skewed every depolarizing/combined-noise circuit heavily toward Z regardless of p (verified: at p=0.3, P(X|fire)=P(Y|fire)=10%, P(Z|fire)=80% instead of the documented 33.3% each — confirmed both via an isolated 100k-sample trace of the raw branch logic and via full statevector simulation, both matching the bug's predicted skew to within statistical noise). Found via independent statistical fuzz testing comparing measured frequencies against the analytic prediction — a test that only checks "the channel runs without crashing" would never have caught this. bitflip, phaseflip, and amplitude_damping were verified unaffected (correct by construction, don't use this three-way branch).
  • Docs: opened a tracking issue for a related robustness gap found during the same fuzzing pass — unrecognized gate names (a typo like 'crx' instead of 'crz', or any gate not in GATE_IDS) are silently dropped everywhere in the simulator instead of raising, same pattern already documented for the Qiskit interop bridge's unsupported custom gates. Not fixed here — would be a breaking-change decision for run_circuit/run_circuit_jit_beast_mode/run_parametric_batch_jit's public behavior, tracked separately.

v8.1.19 — Security fix

  • Fixed (security): QASMParser's gate-parameter evaluator (_eval_param, used for expressions like rx(...), p(...)) called eval() with {'__builtins__': {}} as its only protection. That blocks direct builtin names (open, len, __import__, ...) but does not block attribute/dunder traversal of the live Python object graph (().__class__.__bases__[0].__subclasses__()...), which needs no builtin name at all — from there, any class loaded in the process is reachable, including ones whose __globals__ reference os/subprocess. Verified directly: a crafted gate-parameter expression, passed through the public QASMParser.parse() entry point (the primary entry point of the whole library — used by the dashboard, the Qiskit/PennyLane interop bridge, and any direct usage), executed successfully. Anyone parsing untrusted QASM text was affected, in every previously published version. Fixed by replacing eval() with an AST node-type whitelist evaluator (_eval_ast_node) — only literals, +-*/%** arithmetic, and calls/lookups restricted to the documented math environment (pi, sin, sqrt, ...) are ever evaluated; an ast.Attribute node (produced by any . in the expression) is never one of the handled cases, so attribute-based escapes are structurally impossible rather than blocklisted. _resolve_int_expr (QASM3 for-loop bounds) used the same eval() pattern but was already protected by a pre-filter regex rejecting any non-arithmetic character — verified safe before this fix — now shares the same AST evaluator for consistency, so no raw eval()/exec() remains anywhere in the codebase (confirmed via full-repo search). No public API or behavior change for legitimate expressions — every previously-supported parameter syntax (pi, pi/2, sqrt(2), cos(0.3), etc.) evaluates identically.

If you parse QASM text from any source you don't fully trust, upgrade immediately.

v8.1.18

  • Fixed: removed a global warnings.filterwarnings('ignore') from registry.py, run unconditionally on import dense_evolution. It silenced every Python warning process-wide for the importing user's whole session — not just this package's, but their own code's and every other library's too. Inherited unchanged from the original Colab notebook (added in v8.0.6, never reconsidered once this became a real pip package). Concretely masked real signal: the JAX float64→float32 truncation UserWarnings visible throughout this project's own test output (precision silently lost under use_float32=True) would have been invisible to anyone using the package normally.

v8.1.17

  • Added: donate_argnums=(0,) on run_circuit_jit_beast_mode's statevector buffer — the only one of _compile_and_run_circuit_jit's four call sites where it's safe (self.sv is always rebound immediately after, verified across chunked/repeated calls and separate simulator instances). run_parametric_batch_jit (its init_sv is a vmap-broadcast closure shared across the whole batch) and circuit_to_energy_fn's VQE loop (same stato_zero reused every epoch) are deliberately left un-donated — donating there would make JAX raise on the second use instead of helping. Verified with a real measurement, not just a claim: RSS growth on a 22-qubit/300-gate circuit drops from +89.4MB to +4.5MB.

v8.1.16

  • Note: v8.1.15's published PyPI package does not contain the from_pennylane Python 3.10 fix described below, despite the changelog entry — the fix landed in the repo before the PyPI upload, but the actual pip install-able wheel/sdist for 8.1.15 was built and uploaded from an earlier commit. PyPI doesn't allow re-uploading files under an already-published version, so this release exists specifically to ship that fix as an installable package. If you're on 8.1.15, upgrade to 8.1.16 — don't rely on 8.1.15's changelog matching what you actually have installed.

v8.1.15

  • Added: dense_evolution.autodiff.circuit_to_energy_fn(circuit, n_qubits) — the real VQE gradient engine (jax.value_and_grad through a jax.lax.scan circuit template, verified against finite differences to ~1e-11) is now public API, independent of dashboard_core.py/Streamlit. Takes a QASMCircuit — the same type from_qiskit/from_pennylane return — so it closes the non-differentiability gap documented in v8.1.14: circuit_to_energy_fn(from_pennylane(qnode, ...), n_qubits) now gives a real, non-zero jax.grad, verified directly, where run_pennylane_circuit alone silently returned 0.0.
  • Changed: dashboard_core.py's _build_vqe_template/_vqe_energy_fn removed — _run_vqe_telemetry_body now calls the same public circuit_to_energy_fn, one engine instead of two copies of the same math that could silently drift apart. Verified behaviorally identical: all existing dashboard VQE tests pass unchanged, same tolerances.
  • Fixed: found while testing the newly-public API — calling the engine on a circuit with zero parametric gates crashed on empty-array indexing during JAX tracing. Previously unreachable because dashboard_core.py always special-cased n_params == 0 before calling in; a real gap once this became public API someone could call directly. Fixed with a static (non-traced) branch.
  • Docs: the README's VQE Engine section had drifted stale, still describing the deleted risolvi_qasm() mechanism from before the real-gradient rewrite — corrected, and a new "Differentiable Circuits" section documents circuit_to_energy_fn with a verified end-to-end example.
  • Fixed: from_pennylane broke on Python 3.10 — CI caught it (3.10 job red, 3.11/3.12 green). Newer PennyLane releases dropped Python 3.10 support, so pip resolves an older PennyLane (0.42.3) there instead of the version this bridge was built against (0.45.1); qml.to_openqasm(tape) behaves incompatibly between the two for a bare tape/QuantumScript input (crashes with AttributeError: 'QuantumTape' object has no attribute 'func' on the older one). Verified against both versions directly (installed 0.42.3 in an isolated venv to reproduce). from_pennylane now picks whichever serialization path the installed PennyLane version actually supports instead of assuming the newer one unconditionally.

v8.1.14

  • Added: interop bridge for Qiskit and PennyLane — from_qiskit/from_pennylane convert an existing circuit to a QASMCircuit by reusing the existing QASMParser (via qiskit.qasm2.dumps / qml.to_openqasm) instead of a bespoke gate-by-gate translator; run_qiskit_circuit/run_pennylane_circuit execute it directly on DenseSVSimulator. Handles the bit-order mismatch explicitly instead of leaving it as a silent trap: Qiskit indexes arrays little-endian (qubit 0 = LSB), Dense-Evolution is MSB-first everywhere else in the codebase, so run_qiskit_circuit reorders its output to match Qiskit's own convention (verified against Statevector(...).probabilities() on an asymmetric circuit); PennyLane's own wire order already matches Dense-Evolution's natively (verified the same way), so run_pennylane_circuit does not reorder — kept as two separate code paths on purpose. New optional extras dense-evolution[qiskit] / dense-evolution[pennylane].
  • Fixed: found while building the Qiskit bridge — qiskit.qasm2.dumps exports composite gates (e.g. mcx) as a gate NAME params { ... } definition on a single line, the same brace-delimited block corruption already fixed for QASM3 for/if/while/def in v8.1.13, just not covered because gate wasn't in that fix's keyword set (verified: before the fix, a 4-qubit circuit using mcx silently inflated to n_qubits=5 with a ghost op). Extended the same brace-matching preprocessor to also strip gate definitions cleanly.

v8.1.13

  • Fixed: QASMParser declared OpenQASM 3.0 support but for/if/while/def blocks — brace-delimited, not ;-terminated — were mishandled by the naive split(';') statement splitter: a for-loop's body was never extracted, and its closing } merged into whatever real statement followed on the same line, corrupting it too (verified: for int i in [0:2] { h q[i]; } cx q[0],q[1]; produced a single ghost op named '}', with the loop body lost and the real cx silently dropped — executed circuit stayed |000⟩ at 100% probability, no error). Needed for writing VQE ansätze with a loop over qubits instead of one line per qubit. Added _process_block_constructs, run before the ;-split: for-loops with resolvable integer bounds (literals, or int/const int variables declared earlier in the source — QASM3's inclusive-end range semantics) are now genuinely unrolled by substituting the loop variable into the body per iteration; if/while/def blocks and for-loops with unresolvable bounds are cleanly stripped instead of corrupting the source that follows them.

v8.1.12

  • Fixed: run_circuit_jit_beast_mode / run_parametric_batch_jit silently dropped cy, cp, crz, u1, p, sx — they weren't in GATE_IDS, so if name not in GATE_IDS: continue skipped them with no error (verified: h(0);h(1);crz(0,1,1.2) produced the exact same output as h(0);h(1) alone). dashboard_core.py already treats these as first-class gates, so any circuit using them — dashboard-built or hand-written QASM — silently ran the wrong physics through the fast path nearly everything uses. Added the missing GATE_IDS entries and the missing kernel implementations for cy/crz/sx in _apply_gate_fast_step — crz specifically needed its own kernel, not reuse of cp's (CP phases |11⟩ only; CRZ phases the target conditioned on its own bit, a different gate).
  • Fixed: run_circuit_jit_beast_mode used the raw qubit index as bit position (LSB-first) instead of the documented MSB-first convention (phys = n_qubits - 1 - qubit) used by run_circuit()/apply_gate_1q()/apply_gate_2q()/measure() elsewhere in the simulator. Pre-existing, not introduced by the fix above — found while verifying it, masked until now because every beast-mode circuit tested to date happened to be symmetric under qubit reversal (Bell states, GHZ states, uniform superpositions), so the wrong labeling never showed up in the probabilities. Verified with X on qubit 0 in a 3-qubit register: gave index 1 (LSB) instead of index 4 (MSB, correct). do_1q/do_2q now compute physical bit positions consistently with the rest of the simulator; Chunk's num_chunks==1 (via beast mode) and num_chunks>1 (via apply_gate_1q/apply_gate_2q) paths are now finally consistent with each other too.
  • Fixed: the VQE gradient (run_vqe_telemetry) was never a real derivative — grad_vqe_params[i] = 0.5*(energy-target)*sin(theta[i]) + gaussian_noise, no jax.grad, no parameter-shift rule, no backprop on θ anywhere in the codebase (the only real jax.value_and_grad usage, in QMMMForceEngine, differentiates classical QM/MM forces w.r.t. atomic positions, not circuit parameters). risolvi_qasm (the old circuit-building path) converted θ to a Python float before use, severing the JAX trace, so backprop couldn't pass through it. Replaced with a real jax.grad pipeline reusing run_parametric_batch_jit's own sentinel-injection pattern (θ substituted via jnp.where inside a jax.lax.scan, never a float() call) — verified against a finite-difference gradient (~1.5e-10 agreement) on a real circuit from QASM_LIBRARY, and confirmed genuine Adam-optimizer convergence (monotonic energy descent to a minimum) over 40 epochs, unlike the old noisy formula. Public signature and DataFrame columns of run_vqe_telemetry unchanged.

v8.1.11

  • Fixed: dash.py (the original Colab notebook) was declared as an installable module (py-modules = ["dash"]) with the same name as the real Plotly dash package, itself listed as an optional dependency in the very same pyproject.toml — a genuine packaging collision, not just a local dev annoyance. It also had unconditional module-level from google.colab import files / import ipywidgets, so import dash crashed immediately outside Colab. Nothing in the maintained codebase (dashboard_core.py/app_dashboard.py, the real Streamlit port) imports it anymore. Moved to legacy/dash.py (reference only, not packaged), removed from py-modules. The dashboard extra now installs what the real dashboard actually needs (streamlit, pandas, seaborn, plotly) instead of the unused dash package.
  • Docs: README's Quick Start (the very first example in the file) passed circuit.ops — raw dicts — to run_circuit_jit_beast_mode, which expects the tuple format from circuit.to_tuples(); crashed with KeyError: 0. Fixed, and the "Dashboard" quick-start snippet now points at streamlit run app_dashboard.py instead of the retired Colab-only import dash pattern.

v8.1.10

  • Fixed: run_circuit_jit_beast_mode / run_parametric_batch_jit — a gate referencing a qubit index out of range silently corrupted the entire statevector to zero instead of raising (verified: get_probabilities().sum() went from 1.0 to 0.0, no exception). apply_gate_1q/apply_gate_2q already validated qubit indices, but these two JIT fast paths build their own compiled ops and never called them. Both now validate before dispatch, matching the existing behavior of the non-JIT path.
  • Fixed: Chunk — for n_qubits beyond the RAM-safe budget (chunk_size_bits), it silently ran the circuit on a smaller inner simulator (min(n_qubits, chunk_size_bits)) instead of genuinely chunking: num_chunks/chunk_dim were computed but never used to combine multiple pieces. Found testing Chunk(n_qubits=28): get_probabilities() returned 2**27 elements, not 2**28. Now implements real multi-chunk simulation (RAM-only, no disk paging — covers moderate overflow beyond the safe budget, not arbitrarily large qubit counts): num_chunks independent chunk-sized simulators held in memory, with gate dispatch across chunk boundaries for all six local/chunk-select combinations. Verified against a plain DenseSVSimulator running the identical circuit (exact match, not just "looks right"). A sized RAM check now raises MemoryPressureError up front if the chunks wouldn't fit, instead of attempting and OOMing.

v8.1.9

  • Fixed: ia_utils/vector_healing.py — enhanced_dense_healing_hybrid had an unreachable third branch (a dense/blend fallback): the underlying trigger signal from evaluate_phi_trigger is strictly binary (0.0/1.0), so the branch could never execute. Collapsed to the genuine 2-state logic (pass-through vs. median fallback); runtime output is unchanged since the branch never ran.
  • Fixed: dashboard_core.py — run_simulation / run_vqe_telemetry mutated the process-wide JAX jax_enable_x64 flag without ever restoring it, so running one float32 simulation silently downgraded numerical precision for unrelated code later in the same process (e.g. the Vector Healing page, which sets no precision of its own). Both now save/restore the flag around their own execution.
  • Docs: README's NoiseModel example called a nonexistent .apply() method with a wrong parameter name (n_qubits instead of n) — corrected to apply_to_sv(sv, n=..., ...). Documented QASMCircuit.to_tuples() and DenseSVSimulator.run_circuit, which already existed and work correctly but were never mentioned in the README.

v8.1.8

  • Fixed: parser.py — controlled two-qubit gates (cx/cy/cz/cp/crz) parsed from QASM in the dashboard layer had control and target swapped relative to compiler.py's documented (gate, control, target) contract, breaking entanglement for circuits run through the dashboard. The core QASMCircuit.to_tuples() path was already correct.
  • Fixed: parser.py — range syntax (q[0:3]) on single-qubit gates only applied to the first qubit in the range, silently dropping the rest. Now expands into one gate application per qubit, matching the parser's own documented contract.
  • Fixed: from dense_evolution import Chunk raised ImportError — Chunk is now re-exported from the package root. Added get_probabilities()/get_statevector() to Chunk for parity with DenseSVSimulator.
  • Removed: dense_evolution/test2.py and stress_test.py — byte-identical, assertion-free debug scripts that shipped inside every install with 0% test coverage. Their one real check (Kraus noise is genuinely stochastic across independent runs) is now a real regression test.

v8.1.7

  • ia_utils/ — new package: median_healing, enhanced_dense_healing_hybrid for vector sequence healing (NaN/Inf-safe)
  • jax import in ia_utils.vector_healing made lazy — importable without the [jax] extra
  • Fixed reconstruction_error telemetry returning NaN when input contained NaN/Inf
  • Added scipy to core dependencies (was used but undeclared)

v8.1.6

  • Modular package structure (dense_evolution/ directory)
  • Split registry.py, gates.py, healing.py, chunk.py into dedicated modules

v8.1.5

  • chunk.py — SafeMemoryGuard: hard block at configurable free-RAM threshold (default 15%), soft warning at 2× threshold, gc.collect() before every check
  • chunk.py — Chunk no longer subclasses DenseSVSimulator; inner simulator allocated at safe_qubits only — eliminates RESOURCE_EXHAUSTED on 28q–34q circuits
  • chunk.py — CircuitChunker.split_circuit RAM-checks every gate-slice before dispatch
  • chunk.py — MemoryChunker attributes (num_chunks, chunk_size_bits, dtype) forwarded as @property on Chunk for benchmark compatibility

v8.1.0

  • healing.py — Predictive State Engine: calculate_phi_ab, calculate_vettore_dinamico, calculate_delta_preemp, evaluate_phi_trigger, calculate_jax_reflection — all @jax.jit
  • MemoryReflectionEngine — event logging + JAX Zero-Drift spectral aggregation

v8.0.x

  • run_parametric_batch_jit() — jax.vmap over full parameter grids in single XLA call
  • run_circuit_jit_beast_mode() — static JIT compilation with QuantumTranspiler
  • OpenQASM 2.0/3.0 dual-mode parser with paren-depth-aware expression splitting
  • NoiseModel Kraus channels in registry.py