Given the two 1D numpy arrays a and b with
N = 100000
a = np.randn(N)
b = np.randn(N)
Why is there a considerable execution time difference between the following two expressions:
# expression 1
c = a @ a * b @ b
# expression 2
c = (a @ a) * (b @ b)
Using the %timeit magic of Jupyter Notebook I get the following results:
%timeit a @ a * b @ b
223 µs ± 6.97 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
and
%timeit (a @ a) * (b @ b)
17.4 µs ± 27.3 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
In both versions you do two dot products of length-N vectors. However, in addition the first solution performs N multiplications while the second solution only needs one.
a @ a * b @ b is equivalent to ((a @ a) * b) @ b or
aa = a @ a # N multiplications and additions -> scalar
aab = aa * b # N multiplications -> vector
aabb = aab @ b # N multiplications and additions -> scalar
(a @ a) * (b @ b) is equivalent to
aa = a @ a # N multiplications and additions -> scalar
bb = b @ b # N multiplications and additions -> scalar
aabb = aa * bb # 1 multiplication -> scalar
The fact that matrix multiplication performance can depend on how to set the parentheses is well known. There exist algorithms to optimize matrix chain multiplication by exploiting this fact.
Update: As I just learned, numpy has a function for optimizing multiple matrix multiplications: numpy.linalg.multidot