Skip to content

pruning.md

Latest commit

 

History

History
52 lines (42 loc) · 2.51 KB

pruning.md

File metadata and controls

52 lines (42 loc) · 2.51 KB
Raw

Circuit pruning

Pruning skips gates outside the support reachable from the input observable. It is opt-in and is most useful for local observables and local circuits.

pruning Behavior
None (default) Execute the complete circuit.
"light-cone" Plan once from the initial SPO.
"light-cone-barrier" Refresh from the current SPO before each nonempty barrier-delimited block. 
final_spo, info = spd.evolve(
    initial_spo, circuit, trunc_val, max_num_str,
    pruning="light-cone-barrier", progress=False,
)
print(info["pruning"])  # method and total/retained/pruned gate counts

terminal = spd.init_gradient_spo(final_spo, basis="0")
initial_spgo, gradients, backward_info = spd.backpropagate(
    terminal, circuit, trunc_val, max_num_str, progress=False,
)

Both modes propagate observables in reverse circuit order. Barrier refresh can find tighter support after commutation and truncation. It uses one forward call and one combined record: backward replays exactly the selected forward gates, without deriving a cone from the gradient operator. Use the full original circuit, including the same angles, for backward. Gradient and diagnostic entries retain their original indices; omitted rotations have zero gradients.

Pytket and OpenQASM barriers are supported. Direct IR represents them as SkippedOperation("barrier") or SkippedOperation("OpType.Barrier"). Consecutive barriers and measurement-only blocks add no scans. With no barriers, barrier mode uses one whole-circuit plan. No gates are reordered.

Both modes work on NumPy, JAX and Triton, including Triton's evolve_step. Triton reduces support on GPU, transfers only the packed mask and validity flag, and avoids state copies between blocks. Scans still depend on state size; packed key widths and memory usage still depend on the number of qubits.

Skipped gates also skip truncation and term caps. There is no initial cleanup: a sub-threshold input coefficient can survive if all gates are omitted. Retained gates use the requested numerical policy. The returned backward SPGO describes the reduced circuit; its adjoint coefficients need not equal the unpruned result.

Keep the forward result unmodified until gradient initialization. Another forward call replaces its record, so chaining calls does not assemble a backward record for the concatenated circuit. Use barrier mode within one call for that workflow. Pruned noise analysis is not supported.

See the benchmark note for timings and reproduction.