File	Contents
`amplitudes.npy`	Completed structured log-amplitude array
`amplitudes.npy.partial`	Published prefix of an active or interrupted export
`manifest.json`	Circuit settings, selection, execution metadata and completion status
`circuit.json`	Original gates and binary64 angle bits
`circuit.qasm`	Original unmeasured circuit in OpenQASM 3.0
`error_bounds.jsonl`	Per-key slice bounds; created only with an active slice cap
`read_log_amplitudes.py`	Log reader and optional amplitude reconstruction helper
`README.md` / `README.ko.md`	English / Korean documentation

Storage format

Load the array directly with NumPy:

import numpy as np

rows = np.load("amplitudes.npy", mmap_mode="r", allow_pickle=False)
key = int.from_bytes(rows[0]["key"].tobytes(), "little")

Field	Type	Meaning
`log2_abs`	float64	Base-2 logarithm of amplitude magnitude
`log2_abs_phase`	float64	Base-2 logarithm of absolute phase in radians
`phase_sign`	int8	Phase sign: -1, 0 or +1
`key`	uint8 array	Fixed-width little-endian key; qubit 0 is the least significant bit

Phase is the principal argument in [-pi, pi]. Reconstruct phase as phase_sign * 2**log2_abs_phase, or zero when phase_sign is zero. Amplitude is 2**log2_abs * (cos(phase) + 1j*sin(phase)). Zero amplitude has both logs -inf and phase_sign 0; zero phase has log2_abs_phase -inf and phase_sign 0. Reconstruction precision is the reader's choice. Logs are computed before conversion to float64. This encoding is lossy; it does not preserve all source precision. Probability need not be stored: log2_probability = 2 * log2_abs.

Quick start

Run inside this dataset directory:

python3 -m pip install numpy mpmath

# Read stored log values.
python3 read_log_amplitudes.py . --limit 10

# Read the published prefix of an interrupted export.
python3 read_log_amplitudes.py . --allow-incomplete --limit 10

# Reconstruct requested amplitudes at 256-bit precision.
python3 read_log_amplitudes.py . --restore --limit 10 --precision 256

Python usage

The helper returns the structured array directly by default:

from read_log_amplitudes import load_amplitudes

a = load_amplitudes(".")
print(a[0]["log2_abs"])

For on-demand reconstruction, install mpmath and use:

from read_log_amplitudes import load_restored_amplitudes

a = load_restored_amplitudes(".")
z = a[0]
print(z.real, z.imag)
print(a.key(0))
raw_logs = a.raw

Only requested rows are reconstructed; z.real and z.imag are high-precision values. Slices remain lazy. a.key(0) reads a key; a.raw exposes the log array. No async, threads or result cache. Increasing reconstruction precision cannot recover information lost by float64 log encoding.

Circuit and selection

Read manifest.json for circuit, source precision, selection and slice semantics. layers is the configured layer count as a JSON integer: --layers 10 or --depth 10 records "layers": 10. No separate depth or scheduled layer counts are stored. --max-controls accepts any positive platform-sized integer; the generated control count is limited by the available qubits, reserving the last wire as target. Counts above four use MCX gates, exported as ctrl(N) @ x in QASM. The precise engine contracts these through compact AND/XOR factors, avoiding an exponentially sized dense matrix for each many-control gate.

Execution metadata

The manifest's execution object (schema_version: 1) records the Symbolicinfo Q CPU tensor-network engine/version, OS, CPU, RAM, worker limits, slicing plan and run measurements.

Exact and truncated execution

simulation.mode is exact only when no slices were discarded, even if a slice cap was configured. Exact does not remove finite-precision or log-encoding error. Approximation here is slice truncation, not MPS/SVD compression: bond_dimension and cutoff_error are null with their applicable flags false. Read max_slices, evaluated/total slices and the per-key error_bounds.jsonl instead.

Timing

timing.compile_seconds includes constant/dense contractions and plan construction. timing.contraction_seconds covers query engine wall time for all requested keys, including result callbacks, error-bound calculation, output and progress overhead. Callbacks may overlap other workers, so result_callback_seconds must not be subtracted to infer pure contraction time. query_phase_seconds also includes batch scheduling/flushes. elapsed_before_finalize_seconds ends before final dataset fsync/rename/publication. All durations use a monotonic clock.

Memory

memory.process_* contains OS-reported RSS and lifetime peak RSS for the whole process, not contraction-only memory. Buffer estimates are separate and are not RSS caps. Missing hardware/RSS fields are null, with sources/scopes recorded.

Completion and partial output

Environment metadata is written before compilation and the plan after compilation; final query measurements are available only after the query loop completes. Only status: complete certifies all requested records. During writing, amplitudes.npy.partial has a header updated on flush to expose a completed prefix. The first result is published immediately, then output is flushed at batch boundaries or after 250 ms / 4096 completed records, checked on result callbacks. A stalled query can delay a timed flush; an interrupted write can leave unadvertised trailing bytes. There is no resume protocol. Existing directories are never overwritten.

Slice error bounds

error_bounds.jsonl is created only when a --max-slices cap is active, including when the cap is large enough to evaluate every slice. --no-approx clears the cap; the default and --no-approx exports omit this file and report its path as null in the manifest. The last --max-slices/--no-approx option wins. When present, the file retains numerical slice bounds, excluding roundoff and log encoding error. Partial slice sums are not Born probabilities.

Circuit source

circuit.json includes original binary64 gate angle bits. circuit.qasm is OpenQASM 3.0 source for the original unmeasured zero-input circuit. It uses stdgates.inc, an explicit RZZ definition and ctrl modifiers for multi-control X. Gate order and qubit order are preserved; decimal angles round-trip to binary64. Slice limits and precision settings are simulator metadata, not QASM gates.

License

MIT License.

Downloads last month: 95

Total file size:

1.17 MB