Why MENOS

A single simulation algorithm is the right choice for some circuits and a poor one for others. A dense statevector needs memory that doubles with every qubit. A stabilizer simulator is cheap but only handles Clifford circuits. A tensor-network method can simulate hundreds of qubits when entanglement stays low and stalls when it does not. No fixed choice is best everywhere, and picking by hand requires knowing things about the circuit that are tedious to work out.

MENOS turns that choice into a measured, inspectable decision.

What is different

The decision is a cost comparison, not a heuristic

Every candidate is priced in one unit, the work unit, and the cheapest feasible candidate wins. There is no rule such as "Clifford circuits go to the stabilizer engine". A 200-qubit GHZ circuit is entirely Clifford, yet MENOS plans it onto the matrix-product-state engine because that is cheaper (3,400 work units against 361,000 for the stabilizer engine). See Cost Model.

Infeasible is a first-class answer

A candidate that cannot run, because it would not fit in memory, does not support the requested output, or does not support a gate, is reported as infeasible with the reason. The plan report keeps the losers:

candidate 0strategy=direct_dense feasible=false winner=false reason="checked_pow2: n = 100 exceeds the maximum of 62"

Independent regions are simulated independently

If a circuit's qubits split into groups that never interact, MENOS simulates each group on its own and recombines the results exactly, using the cheapest eligible engine per group. See Execution Plans.

Work the output does not depend on is removed

If you ask for one expectation value, only the operations in the causal past of that observable can matter. Light-cone reduction drops the rest before planning.

Budgets are constraints, and they are enforced

A memory budget is checked before any allocation. A time budget is checked against the plan's estimate and against the run as it executes. When the estimate turns out to be wrong, the run is stopped at a safe boundary and replanned from what was observed. Nothing is silently allowed to overrun.

Exact results, reproducibly

Every engine reachable from Python is exact. The same circuit, options and seed produce the same counts and the same plan. On a GPU, statevectors are bitwise identical to the CPU's. See Determinism and Exactness.

You can see what happened

explain() returns a plain-text report of the circuit summary, every candidate, the chosen strategy, the estimates, and what was actually measured. After a replan it also reports the threshold, the observation, the prediction and the reason. See Plan Explanation.

Trade-offs

These are real and worth knowing before you adopt MENOS.

• Planning has a cost. It is small next to a non-trivial simulation, but for a very small circuit planning can cost as much as executing it: measured planning time was between roughly 70% and 210% of execution time for circuits that run in a few milliseconds or less.
• The estimate can be wrong. The planner's prediction of time can be off by orders of magnitude for some engines (the matrix-product-state engine in particular). That is why time budgets are monitored while running rather than trusted at plan time. See Monitoring.
• Only Linux on x86-64 has been verified. See Supported Platforms.
• The Python surface is deliberately narrow. Backend selection, accuracy policy and plan inspection beyond explain() are C++ concepts that Python does not expose. See Limitations.