Choose a Contraction Path Finder

Goal

Configure how the tensor network contraction order is determined. The contraction path finder strongly affects performance and peak memory of tensor-network-based estimators and samplers.

Overview

Both the tensor network estimator and the tensor network sampler expose a choose_contraction_path_finder() method. The available path finders are:

.sequential() (default): tries a fast greedy heuristic first, then falls through to the comprehensive multi-strategy finder if the greedy path exceeds resource thresholds.
.random_greedy_128(): randomized greedy algorithm with 128 trials. A quick heuristic.
.multi_strategy(): runs multiple strategies and returns the best path.
.none(): Choose no contraction path finder. This will leave finding the contraction path up to the contraction calculator.

If choose_contraction_path_finder() is not called, the default sequential() path finder is used.

Sequential Path Finder (default)

The sequential path finder tries path finders in order (cheapest first) and returns the first path that satisfies both a FLOPs and a memory threshold. By default it uses:

Tier 1 (greedy): fast heuristic, accepted for small problems.
Tier 2 (multi-strategy): comprehensive search across multiple strategies, always accepted as the final fallback.

This code example is for estimator - but it is the same for sampler.

import qrunch as qc

# Use default sequential (greedy -> multi-strategy)
estimator = (
      qc.estimator_creator()
      .tensor_network()
      .choose_contraction_path_finder()
      .sequential()
      .create()
)

Greedy algorithm with 128 trials

Randomized greedy algorithm with 128 trials. A quick heuristic. This code example is for estimator - but it is the same for sampler.

import qrunch as qc

# Use default sequential (greedy -> multi-strategy)
estimator = (
      qc.estimator_creator()
      .tensor_network()
      .choose_contraction_path_finder()
      .random_greedy_128()
      .create()
)

Multi-Strategy Path Finder

The .multi_strategy() path finder runs multiple strategies and returns the best path (lowest FLOPs) that fits within the memory budget. Falls back to slicing/memory-limited strategies if no path fits.

It does not include a fast greedy heuristic — use sequential() if you want greedy as a fast first tier before the multi-strategy search. This code example is for estimator - but it is the same for sampler.

import qrunch as qc

estimator = (
      qc.estimator_creator()
      .tensor_network()
      .choose_contraction_path_finder()
      .multi_strategy(max_memory_fraction=0.8)
      .create()
)