带批量更新的方差/标准的韦尔福德算法的公式是什么？

Question

我想扩展 Welford 的在线算法，以便能够使用多个数字（批量）进行更新，而不是一次只更新一个： https: //en.wikipedia.org/wiki/Algorithms_for_calculate_variance

我尝试从 wiki 页面更新算法，如下所示：

# my attempt.
def update1(existingAggregate, newValues):
    (count, mean, M2) = existingAggregate
    count += len(newValues) 
    delta = np.sum(np.subtract(newValues, [mean] * len(newValues)))
    mean += delta / count
    delta2 = np.sum(np.subtract(newValues, [mean] * len(newValues)))
    M2 += delta * delta2

    return (count, mean, M2)

# The original two functions from wikipedia.
def update(existingAggregate, newValue):
    (count, mean, M2) = existingAggregate
    count += 1 
    delta = newValue - mean
    mean += delta / count
    delta2 = newValue - mean
    M2 += delta * delta2

def finalize(existingAggregate):
    (count, mean, M2) = existingAggregate
    (mean, variance, sampleVariance) = (mean, M2/count, M2/(count - 1)) 
    if count < 2:
        return float('nan')
    else:
        return (mean, variance, sampleVariance)

但是，我一定理解不正确，因为结果是错误的：

# example x that might have led to an a = (2, 2.0, 2.0).
x = [1.0, 3.0]
mean = np.mean(x)
count = len(x)
m2 = np.sum(np.subtract(x, [mean] * count)**2)

a = (count, mean, m2)
print(a)
# new batch of values.
b = [5, 3]

请注意，a = (2, 2.0, 2.0) 表示我们有 2 个观测值，它们的平均值为 2.0。

# update one at a time.
temp = update(a, newValues[0])
result_single = update(temp, newValues[1])
print(finalize(result_single))

# update with my faulty batch function.
result_batch = update1(a, newValues)
print(finalize(result_batch))

正确的输出应该是应用单个数字更新两次的输出：

(3.0, 2.0, 2.6666666666666665)
(3.0, 2.5, 3.3333333333333335)

关于正确的方差更新，我缺少什么？我是否还需要以某种方式更新 Finalize 函数？

我需要这样做的原因是因为我正在处理非常大的每月文件（具有不同数量的观察结果）并且我需要获得每年的平均值和方差。

Answer 1

感谢尼科的澄清，我明白了！问题是我对增量求和，然后乘以得到 M2，但必须对增量的乘积求和。这是正确的批处理函数，它能够接受单个数字以及批量：

# https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
def update(existingAggregate, newValues):
    if isinstance(newValues, (int, float, complex)):
        # Handle single digits.
        newValues = [newValues]

    (count, mean, M2) = existingAggregate
    count += len(newValues) 
    # newvalues - oldMean
    delta = np.subtract(newValues, [mean] * len(newValues))
    mean += np.sum(delta / count)
    # newvalues - newMeant
    delta2 = np.subtract(newValues, [mean] * len(newValues))
    M2 += np.sum(delta * delta2)

    return (count, mean, M2)

def finalize(existingAggregate):
    (count, mean, M2) = existingAggregate
    (mean, variance, sampleVariance) = (mean, M2/count, M2/(count - 1)) 
    if count < 2:
        return float('nan')
    else:
        return (mean, variance, sampleVariance)

用法示例：

x = [1.0, 3.0]
mean = np.mean(x)
count = len(x)
m2 = np.sum(np.subtract(x, [mean] * count)**2)

a = (count, mean, m2)
print(a)
# new batch of values.
b = [5, 3]

result_batch = update(a, b)
result_batch1 = update(a, b[0])

print(finalize(result_batch))
print(finalize(result_batch1))

而且它确实更快：

import timeit
x = random.sample(range(1, 10000), 1000)
# ...
b = random.sample(range(1, 10000), 1000)

start_time = timeit.default_timer()
result_batch = update(a, b)
print(f'{timeit.default_timer() - start_time:.4f}')
print(*(f'{x:.2f}' for x in finalize(result_batch)))

start_time = timeit.default_timer()
for i in b:
    a  = update1(a, i)
print(f'{timeit.default_timer() - start_time:.4f}')
print(*(f'{x:.2f}' for x in finalize(result_batch)))

结果：

0.0010
5008.36 8423224.68 8427438.40
0.0031
5008.36 8423224.68 8427438.40