2.1 `random_walk_fastest` is slightly incorrect: does not include starting position `0`

There is a subtle incorrectness in random_walk_fastest. It does not return the first position 0, compared to other two implementations. The following code demonstrates this:

N = 99
print(len(random_walk(N)), len(random_walk_faster(N)), len(random_walk_fastest(N)))
# >>> (100, 100, 99)

This may look as a minor distinction, but may cause bugs. There are two ways to remove this inconsistency: exclude initial 0 position from first two functions (trivial) or prepend it to the last function. There are such options for the latter path:

Convert steps to numpy list after cumsum and concat with [0]
Concatenate np.array([0])
Allocate empty array in advance, and use out parameter of cumsum

My implementation of these options:


N = 10_000

# Option 1: Convert `steps` to python list after `cumsum` and concat with `[0]`
def random_walk_option1(n=N):
    steps = np.random.choice([-1,+1], n)
    steps = np.cumsum(steps)
    return [0] + steps.tolist()

# Option 2.1: Concatenate `np.array([0])`
def random_walk_option2_1(n=N):
    steps = np.random.choice([-1,+1], n)
    steps = np.cumsum(steps)
    return np.concatenate((np.array([0]), steps))

# Option 2.2: Concatenate `[0]`
def random_walk_option2_2(n=N):
    steps = np.random.choice([-1,+1], n)
    steps = np.cumsum(steps)
    return np.concatenate(([0], steps))

# Option 3: Allocate empty array in advance, and use `out` parameter of `cumsum`
def random_walk_option3_1(n=N):
    walk = np.empty(n+1)
    walk[0] = 0
    steps = np.random.choice([-1,+1], n)
    np.cumsum(steps, out=walk[1:])
    return walk

def random_walk_option3_2(n=N):
    walk = np.zeros(n+1)
    steps = np.random.choice([-1,+1], n)
    np.cumsum(steps, out=walk[1:])
    return walk

I also benchmarked them to see the difference: https://colab.research.google.com/drive/19OJdrD4SZLk4ug6OI6DC6wq3XwHQPAgr?usp=sharing

While option_2_1 is slightly better, the benchmark is not stable. Maybe it is because the difference in the number of operations is up to the constant, the difference is not significant?

Fixing it in a book may be nasty/non-trivial because will break the simplicity of implementation and add more confusion for the reader :)

rougier / from-python-to-numpy

2.1 `random_walk_fastest` is slightly incorrect: does not include starting position `0` #119