accurate sums and products for Python
Accurate sums and (dot) products for Python.
Summing up values in a list can get tricky if the values are floating point numbers; digit cancellation can occur and the result may come out wrong. A classical example is the sum
1.0e16 + 1.0 - 1.0e16
The actual result is 1.0, but in double precision, this will result in 0.0. While in this example the failure is quite obvious, it can get a lot more tricky than that. accupy provides
p, exact, cond = accupy.generate_ill_conditioned_sum(100, 1.0e20)
which given a length and a target condition number will produce an array if floating point numbers that’s hard to sum up.
accupy has the following methods for summation:
- accupy.kahan_sum(p): Kahan summation
- accupy.fsum(p): A vectorization wrapper around math.fsum (which uses Shewchuck’s algorithm  (see also here)).
- accupy.ksum(p, K=2): Summation in K-fold precision (from )
All summation methods sum the first dimension of a multidimensional NumPy array.
Let’s compare them.
Accuracy comparison (sum)
As expected, the naive sum performs very badly with ill-conditioned sums; likewise for `numpy.sum <https://docs.scipy.org/doc/numpy/reference/generated/numpy.sum.html>`__ which uses pairwise summation. Kahan summation not significantly better; this, too, is expected.
Computing the sum with 2-fold accuracy in accupy.ksum gives the correct result if the condition is at most in the range of machine precision; further increasing K helps with worse conditions.
Shewchuck’s algorithm in math.fsum always gives the correct result to full floating point precision.
Speed comparison (sum)
We compare more and more sums of fixed size (above) and larger and larger sums, but a fixed number of them (below). In both cases, the least accurate method is the fastest (numpy.sum), and the most accurate the slowest (accupy.fsum).
accupy has the following methods for dot products:
- accupy.fdot(p): A transformation of the dot product of length n into a sum of length 2n, computed with math.fsum
- accupy.kdot(p, K=2): Dot product in K-fold precision (from )
Let’s compare them.
Accuracy comparison (dot)
accupy can construct ill-conditioned dot products with
x, y, exact, cond = accupy.generate_ill_conditioned_dot_product(100, 1.0e20)
With this, the accuracy of the different methods is compared.
As for sums, numpy.dot is the least accurate, followed by instanced of kdot. fdot is provably accurate up into the last digit
Speed comparison (dot)
NumPy’s numpy.dot is much faster than all alternatives provided by accupy. This is because the bookkeeping of truncation errors takes more steps, but mostly because of NumPy’s highly optimized dot implementation.
accupy needs the C++ Eigen library, provided in Debian/Ubuntu by `libeigen3-dev <https://packages.ubuntu.com/search?keywords=libeigen3-dev>`__.
accupy is available from the Python Package Index, so with
pip install -U accupy
you can install/upgrade.
To run the tests, just check out this repository and type
To create a new release
bump the __version__ number,
publish to PyPi and GitHub:
$ make publish
accupy is published under the MIT license.