def test_gauss_suff_stats():
# High mean, tiny variance would lead to catastrophic cancellation
# in a naive implementation that maintained the sum of squares.
big = 400
small = 0.0000001
data = [big - small, big, big + small]
true_sigma = math.sqrt(2 * small**2 / 3)
(ct, mean, sigma) = stats.gauss_suff_stats(data)
assert ct == 3
assert mean == big
assert relerr(true_sigma, sigma) < 1e-5
def estimate_mean(samples):
"""Estimate the mean of a distribution from samples.
Return the triple (count, mean, error).
`count` is the number of input samples.
`mean` is the mean of the samples, which estimates the true mean
of the distribution.
`error` is an estimate of the standard deviation of the returned
`mean`. This is computed from the variance of the input samples,
on the assumption that the Central Limit Theorem applies. This is
will be so if the underlying distribution has a finite variance,
and enough samples were drawn.
"""
(n, mean, stddev) = stats.gauss_suff_stats(samples)
return (n, mean, stddev / math.sqrt(n))
def estimate_mean(samples):
"""Estimate the mean of a distribution from samples.
Return the triple (count, mean, error).
`count` is the number of input samples.
`mean` is the mean of the samples, which estimates the true mean
of the distribution.
`error` is an estimate of the standard deviation of the returned
`mean`. This is computed from the variance of the input samples,
on the assumption that the Central Limit Theorem applies. This is
will be so if the underlying distribution has a finite variance,
and enough samples were drawn.
"""
(n, mean, stddev) = stats.gauss_suff_stats(samples)
return (n, mean, stddev / math.sqrt(n))