#!/usr/bin/env python3
# Source: t5-linear-algebra.html
import numpy as np

def softmax(z):
    e = np.exp(z - z.max())          # shift for stability
    return e / e.sum()

rng = np.random.default_rng(3)
z = rng.normal(size=4)
p = softmax(z)
J = np.diag(p) - np.outer(p, p)      # analytic Jacobian

h = 1e-6
I = np.eye(4)
J_num = np.column_stack([
    (softmax(z + h * I[i]) - softmax(z - h * I[i])) / (2 * h)
    for i in range(4)
])
print(np.allclose(J, J_num, atol=1e-8))   # True
print(np.round(J.sum(axis=0), 12))        # columns sum to 0
