[Numpy-discussion] Tensor contraction
Friedrich Romstedt
friedrichromstedt@gmail....
Sun Jun 13 14:37:50 CDT 2010
2010/6/13 Alan Bromborsky <abrombo@verizon.net>:
> I am writing symbolic tensor package for general relativity. In making
> symbolic tensors concrete
> I generate numpy arrays stuffed with sympy functions and symbols.
That sound's interesting.
> The
> operations are tensor product
> (numpy.multiply.outer), permutation of indices (swapaxes), partial and
> covariant (both vector operators that
> increase array dimensions by one) differentiation, and contraction.
I would like to know more precisely what this differentiations do, and
how it comes that they add an index to the tensor.
> I think I need to do the contraction last
> to make sure everything comes out correctly. Thus in many cases I would
> be performing multiple contractions
> on the tensor resulting from all the other operations.
Hm, ok, so I guess I shall give my 1 cent now.
Ok.
# First attempt (FYI, failed):
# The general procedure is, to extract a multi-dimensional diagonal array.
# The sum \sum_{ij = 0}^{M} \sum_{kl = 0}^{N} is actually the sum over a
# 2D array with indices I \equiv i \equiv j and K \equiv k \equiv
l. Meaning:
# \sum_{(I, K) = (0, 0)}^{(M, N)}.
# Thus, if we extract the indices with 2D arrays [[0], [1], ...,
[N - 1]] for I and
# [[0, 1, ..., M - 1]] on the other side for K, then numpy's broadcasting
# mechanism will broadcast them to the same shape, yielding (N, M) arrays.
# Then finally we sum over this X last dimensions when there were X
# contractions, and we're done.
# Hmmm, when looking closer at the problem, it seems that this isn't
# adequate. Because we would have to insert open slices, but cannot
# create them outside of the [] operator ...
# So now follows second attemt:
def contract(arr, *contractions):
"""*CONTRACTIONS is e.g.:
(0, 1), (2, 3)
meaning two contractions, one of 0 & 1, and one of 2 & 2,
but also:
(0, 1, 2),
is allowed, meaning contract 0 & 1 & 2."""
# First, we check if we can contract using the *contractions* given ...
for contraction in contractions:
# Extract the dimensions used.
dimensions = numpy.asarray(arr.shape)[list(contraction)]
# Check if they are all the same.
dimensionsdiff = dimensions - dimensions[0]
if numpy.abs(dimensionsdiff).sum() != 0:
raise ValueError('Contracted indices must be of same dimension.')
# So now, we can contract.
#
# First, pull the contracted dimensions all to the front ...
# The names of the indices.
names = range(arr.ndim)
# Pull all of the contractions.
names_pulled = []
for contraction in contractions:
names_pulled = names_pulled + list(contraction)
# Remove the pulled index names from the pool:
for used_index in contraction:
# Some more sanity check
if used_index not in names:
raise ValueError('Each index can only be used in one
contraction.')
names.remove(used_index)
# Concatenate the pulled indices and the left-over indices.
names_final = names_pulled + names
# Perform the swap.
arr = arr.transpose(names_final)
# Perform the contractions ...
for contraction in contractions:
# The respective indices are now, since we pulled them, the
frontmost indices:
ncontraction = len(contraction)
# The index array:
# shape[0] = shape[1] = ... = shape[ncontraction - 1]
I = numpy.arange(0, arr.shape[0])
# Perform contraction:
index = [I] * ncontraction
arr = arr[tuple(index)].sum(axis=0)
# If I made no mistake, we should be done now.
return arr
Ok, it didn't get much shorter than Pauli's solution, so you decide ...
> One question to
> ask would be considering that I am stuffing
> the arrays with symbolic objects and all the operations on the objects
> would be done using the sympy modules,
> would using numpy operations to perform the contractions really save any
> time over just doing the contraction in
> python code with a numpy array.
I don't know anything about sympy. I think there's some typo around:
I guess you mean creating some /sympy/ array and doing the operations
using that instead of using a numpy array having sympy dtype=object
content?
Friedrich
More information about the NumPy-Discussion
mailing list