SynthesisConfig

class SynthesisConfig(tolerance_error=1e-08, layout_method='trivial', optimizer='bfgs2', global_search='basin_hopping', n_trials_warm=1, n_trials_cold=1, warm_start_from_level=1, refine_best_trial=False, parallel_mode='openmp', use_float=False, round_single_qubit_rotations=None, seed=None, verbosity='silent', barrier=False)

Represent the parameters every quasarcpp synthesis strategy reads.

This dataclass carries only the options that reach all of the synthesis strategies. Anything a single strategy consumes is declared on that strategy’s own config, so a field is never offered where the backend would ignore it.

Variables:

  • tolerance_error – Target compilation error: an average gate infidelity against a unitary target, a state infidelity against a state, the numbers compilation_error() and state_preparation_error() report. Synthesis is considered successful once the compilation error falls below this value, so a small enough tolerance_error makes the compilation numerically exact.

  • layout_method – How logical qubits map to physical qubits at the circuit’s input and at its output. The mappings are fixed, or searched by synthesizing candidates and ranking them: by the overall error under an error-aware strategy, by what layout_objective chooses under the other two. Values: "trivial" (the default) keeps logical qubit \(i\) on physical qubit \(i\) at both ends; "full_search" tries every (initial, final) pair, \((n!)^2\) syntheses on \(n\) qubits; "greedy_search" picks the input mapping, then the output mapping, \(2\,n!\) syntheses; "symmetric_search" covers the same pairs as "full_search" in fewer syntheses, \(n \cdot n!\) on a star; "diagonal_search" tries only pairs where the output mapping equals the input mapping, \(n\) syntheses on a star; or an explicit (initial_permutation, final_permutation) tuple. Synonyms: none/off -> trivial; full/exhaustive -> full_search; greedy/sequential -> greedy_search; diagonal -> diagonal_search. With a readout target (a trailing PauliProductMeasurement), a search also covers the readout transformation. The chosen mappings are recorded on the compiled circuit’s .layout. Same purpose as QiskitConfig’s layout_method, but a different mechanism and value set.

  • optimizer – Name of the backend numerical optimizer used to fit the ansatz parameters, case-insensitive: "bfgs2" (the default), "bfgs", "grad_descend", "agents", "agents_combined" or "adam". Any other name raises ConfigError when the config is built. The three fastest and most accurate are bfgs2, bfgs and grad_descend; agents and adam stop well short of the target error, and adam is two orders of magnitude slower.

  • global_search – Name of the global search that wraps the local optimizer, read only by the BFGS2 optimizer; any other optimizer ignores it. "basin_hopping" (the default) selects a basin-hopping restart loop, "dual_annealing" a dual-annealing one, and "differential_evolution" a differential-evolution one.

  • n_trials_warm – Number of optimizations the backend runs on each candidate gate structure starting from the previous level’s result. The first level has no previous result, and its warm trials start from a random point.

  • n_trials_cold – Number of optimizations the backend runs on each candidate gate structure from a random starting point. The best of all trials, warm and cold, decides the structure’s result, and at least one trial must run. More trials take longer, and can reach tolerance_error on targets that fewer trials miss.

  • warm_start_from_level – First level whose warm trials start from the previous level’s result. Below it every trial starts from a random point, whatever n_trials_warm says. 1 (the default) lets every level start from its predecessor.

  • refine_best_trial – Whether one further optimization refines the best trial, for one more than the trial counts in total. It is off by default: it adds compile time and seldom changes the gate count.

  • parallel_mode – Name of the backend parallelization mode: "sequential" runs on one thread, "openmp" (the default) and "tbb" run on several threads.

  • use_float – Whether to run the quasarcpp synthesis backend in single-precision (float32) instead of the default double-precision (float64). In single precision a tolerance_error below about 1e-6 is not reached reliably.

  • round_single_qubit_rotations – The angles that single-qubit rotation angles are rounded onto once the search has converged, such as (), (0,) or (0, np.pi / 2). Each angle is read modulo \(2\pi\) and stands for its negative too. None (the default) leaves it to the gate set: FixedCROTGateSet rounds onto (0, np.pi / 2) and every other gate set of this package onto (0,). A rotation within 0.5 rad of a target is rounded onto the nearest one: a zero rotation is no gate, a \(\pi\) rotation is an \(X\), and a rotation onto any other target turns by exactly that angle. A rounding is kept while the compilation error stays within tolerance_error, so the compilation error can rise up to it; an error-aware strategy keeps a rounding while the overall error does not rise. () turns rounding off where compile time matters more than the pulse count. Applies to both unitary and state compilation.

  • seed – Seed for the backend’s random number generator: a non-negative Python or NumPy integer, or None to draw one from the OS. Unseeded, repeated compiles of the same circuit return different (though all tolerance-respecting) results. A fixed seed pins the compiled circuit down to its gate structure and every rotation angle, on every strategy and at every other setting, including each parallel_mode.

  • verbosity – How much the compiler prints: "silent" (the default) prints nothing, then "summary", "detailed" and "debug" each add the backend messages of the next level up.

  • barrier – Whether to insert a barrier() after every native entangler in the compiled circuit, for readability when drawing. A barrier separates each layer (single-qubit rotations plus an entangler) from the next, and the last layer from the closing single-qubit layer. It has no effect on synthesis, the native gate count or .metadata.

Raises:

ConfigError – If a named choice is unknown, a trial count or tolerance_error is negative, seed is not a non-negative integer, or round_single_qubit_rotations holds something other than a finite angle, or one angle twice.