Wolfram/TensorNetworks | Paclet Repository

Wolfram`TensorNetworks`

GreedyContractionPath

GreedyContractionPath [tn]
	finds a contraction path through the TensorNetwork tn using a greedy cost-minimization heuristic.

GreedyContractionPath [graph]
	finds a path for a TensorNetwork Graph .

GreedyContractionPath ["Dimensions"dims,"Indices"inds,"Contractions"cs]
	finds a path from explicit network data in association form.

GreedyContractionPath [expr]
	finds a path for an Inactive TensorContract of a TensorProduct .

GreedyContractionPath [inputs,output,sizeDict]
	low-level form taking explicit per-tensor index lists, the output indices, and a size-per-index Association .

Details and Options

▪

returns the same canonical path format as

: a

of integer pairs

{i,j}

in opt_einsum convention. At each step the two tensors at positions

and

are removed and the contraction result is appended at the end of the list.

▪

The greedy search is

(n)

per step in the network size, so it scales to hundreds of tensors where

is intractable. Paths are not provably optimal but in practice are within a small constant factor.

▪

The search is backed by the Rust library function

"optimize_greedy"

from

, loaded on first use through

ExtensionCargo

▪

Setting

"Temperature"

to a numeric value enables stochastic local search; combined with

"RandomSeed"

this produces deterministic-but-randomized paths, useful for benchmarking ensembles.

▪

The following options can be given:

"MemoryWeight"	None	weight given to intermediate-tensor memory cost vs. FLOPs
"Temperature"	None	stochastic-search temperature; higher values explore more candidates
"MaxNeighbors"	None	beam width: number of candidate pairs evaluated at each step
"RandomSeed"	None	RNG seed for reproducible randomized paths
"PreSimplify"	None	fuse duplicate indices and trivial dimensions before search
"FixedIndexing"	None	use SSA index allocation in the search