Exactness
MENOS never silently approximates. A result is exact, or it is flagged, or the job is refused.
What "exact" means here
Exact means no algorithmic approximation: no truncation, no sampling of an expectation value, no discarded singular values. For a noisy job it means exact for the noise model you asked for: the noise is part of the question, and the density-matrix engine answers that question without approximating it. Arithmetic is still double precision, so results agree with Qiskit's reference simulators to about 1e-16, not to infinite precision. The examples in this documentation check that agreement.
Per engine
| Engine | Exactness |
|---|---|
| Dense | Exact |
| Stabilizer | Exact |
| Pauli propagation | Exact; refuses a job whose term count cannot be represented, rather than truncating |
| MPS | Exact by default: feasible only if a rigorous bound on the bond dimension fits the budget |
| Tensor network | Exact; contraction performs no truncation |
| Cut and recombine | Exact; every branch is enumerated, nothing is sampled |
| Factorized | Exact, as its regions are independent |
| Density matrix | Exact for the requested noise model; noise channels and readout errors are applied exactly, nothing is sampled except the requested shots |
| Trajectory | Not exact: noise is sampled. Used only when you opt in; the result is flagged exact = False and an expectation value carries its measured standard error. See Trajectory |
Where Result.exact comes from
Result.exact is the post-execution truth: it is True only if no approximation was actually made. A plan carries a pre-execution estimate of the same thing.
from menos import Circuit, Opcode, RunOptions, run
circuit = Circuit(2, 2)
circuit.append(Opcode.H, [0])
circuit.append(Opcode.CX, [0, 1])
circuit.append_measure(0, 0)
circuit.append_measure(1, 1)
options = RunOptions()
options.shots = 10
options.seed = 1
result = run(circuit, options)
assert result.exact
assert "(exact)" in result.explain()Approximation exists in the C++ core, not in Python
The C++ core has an accuracy policy, Controlled(tolerance), that lets the MPS engine discard singular values below a per-step cutoff. When it truncates, the result is marked approximate, and MENOS reports an a-posteriori error bound built from what was actually discarded (the tolerance bounds each step, not their sum, so the reported bound is the honest figure).
None of that is reachable from Python. There is no way to request it. The only non-exact results Python can produce are trajectory results, which you request explicitly with expectation_precision or NoiseMethod.Auto, and which are flagged exact = False.
Refusal instead of degradation
When an exact run does not fit the constraints, MENOS refuses. It does not fall back to a cheaper, less accurate method:
- an MPS run whose exact bound exceeds the memory budget is infeasible, not truncated;
- a cut plan refuses any non-exact policy;
- a run that overran its time budget and has no other plan throws
ResourceExceeded.