ncon is a Python 3 package that implements the NCon function as described here: https://arxiv.org/abs/1402.0939 This Python implementation lacks some of the fancier features described in the paper, but the interface is the same.
ncon requires numpy and works with numpy ndarrays. It also works with the various tensors from this package, but does not require it.
pip install --user ncon
The only thing this package exports is the function ncon. It takes a list of
tensors to be contracted, and a list index lists that specify what gets
contracted with that. It returns a single tensor, that is the result of the
contraction. Here's how the syntax works:
ncon(L, v, order=None, forder=None, check_indices=True):
The first argument L is a list of tensors.
The second argument v is a list of list, one for each tensor in L.
Each v[i] consists of integers, each of which labels an index of L[i].
Positive labels mark indices which are to be contracted (summed over).
So if for instance v[m][i] == 2 and v[n][j] == 2, then the ith index of
L[m] and the jth index of L[n] are to be identified and summed over.
Negative labels mark indices which are to remain free (uncontracted).
The keyword argument order is a list of all the positive labels, which
specifies the order in which the pair-wise tensor contractions are to be done.
By default it is sorted(all-positive-numbers-in-v), so for instance
[1,2,...]. Note that whenever an index joining two tensors is about to be
contracted together, ncon contracts at the same time all indices connecting
these two tensors, even if some of them only come up later in order.
Correspondingly forder specifies the order to which the remaining free
indices are to be permuted. By default it is
sorted(all-negative-numbers-in-v, reverse=True),
meaning for instance [-1,-2,...].
If both order and forder are provided by the user, then objects other than
integers can be used to label the indices. This has been tested with string
labels, but in principle many other types of objects should work too.
If check_indices=True (the default) then checks are performed to make sure
the contraction is well-defined. If not, an ValueError with a helpful
description of what went wrong is provided.
If the syntax sounds a lot like Einstein summation, as implemented for example
by np.einsum, then that's because it is. The benefits of ncon are that many
tensor networkers are used to its syntax, and it is easy to dynamically
generate index lists and contractions.
Here are a few examples, straight from the test file.
A matrix product:
from ncon import ncon
a = np.random.randn(3, 4)
b = np.random.randn(4, 5)
ab_ncon = ncon([a, b], ((-1, 1), (1, -2)))
ab_np = np.dot(a, b)
assert np.allclose(ab_ncon, ab_np)
Here the last index of a and the first index of b are contracted.
The result is a tensor with two free indices, labeled by -1 and -2.
The one labeled with -1 becomes the first index of the result. If we gave the
additional argument forder=[-2,-1] the transpose would be returned instead.
A more complicated example:
a = np.random.randn(3, 4, 5)
b = np.random.randn(5, 3, 6, 7, 6)
c = np.random.randn(7, 2)
d = np.random.randn(8)
e = np.random.randn(8, 9)
result_ncon = ncon(
(a, b, c, d, e), ([3, -2, 2], [2, 3, 1, 4, 1], [4, -1], [5], [5, -3])
)
result_np = np.einsum("ijk,kilml,mh,q,qp->hjp", a, b, c, d, e)
assert np.allclose(result_ncon, result_np)
Notice that the network here is disconnected, d and e are not contracted
with any of the others. When contracting disconnected networks, the connected
parts are always contracted first, and their tensor product is taken at the
end. Traces are also okay, like here on two indices of c. By default, the
contractions are done in the order [1,2,3,4,5]. This may not be the optimal
choice, in which case we should specify a better contraction order as a keyword
argument.