您要计算的值小于可以使用64位浮点值表示的值。你在评论中给出的一个例子是k = 5007, M = 45956, n = 18969, N = 5267
。对于M
,n
和N
那些值时,下溢PMF为0时k
参数是3478:
In [46]: k = 5007
In [47]: M = 45956
In [48]: n = 18969
In [49]: N = 5267
In [50]: hypergeom.pmf(3476, M, n, N)
Out[50]: 9.8813129168249309e-324
In [51]: hypergeom.pmf(3477, M, n, N)
Out[51]: 4.9406564584124654e-324
In [52]: hypergeom.pmf(3478, M, n, N)
Out[52]: 0.0
的标准方法来解决这一问题是与概率的对数工作。该SciPy的离散分布具备的功能logpmf
和logsf
此:
In [53]: hypergeom.logpmf(3476, M, n, N)
Out[53]: -743.80749253381509
In [54]: hypergeom.logpmf(3477, M, n, N)
Out[54]: -744.95722489454783
In [55]: hypergeom.logpmf(3478, M, n, N)
Out[55]: -746.10790755529888
In [56]: hypergeom.logpmf(5007, M, n, N)
Out[56]: -3952.1782915849763
为了计算hypergeom.sf(k, M, n, N) + hypergeom.pmf(k, M, n, N)
,您可以使用numpy.logaddexp
:
In [58]: np.logaddexp(hypergeom.logsf(k, M, n, N), hypergeom.logpmf(k, M, n, N))
Out[58]: -3952.1508002445375
唯一不方便的是,进一步的计算和比较,必须立足于概率的对数。如果这不适用于您,则必须切换到提供更高精度浮点计算的库(例如mpmath
)。例如,以下功能使用mpmath
计算PMF和生存函数:
def hypergeom_pmf(k, M, n, N):
tot, good = M, n
bad = tot - good
pmf = (mpmath.beta(good+1, 1) * mpmath.beta(bad+1,1) * mpmath.beta(tot-N+1, N+1)/
(mpmath.beta(k+1, good-k+1) * mpmath.beta(N-k+1,bad-N+k+1) * mpmath.beta(tot+1, 1)))
return pmf
def hypergeom_sf(k, M, n, N):
sf = (mpmath.binomial(N, k+1) * mpmath.binomial(M-N, n - k - 1)/mpmath.binomial(M, n) *
mpmath.hyp3f2(1, k + 1 - n, k + 1 - N, k + 2, M + k + 2 - n - N, 1))
return sf
(在hypergeom_pmf(k, M, n, N)
使用的表达式scipy.stats.hypergeom._logpmf
从SciPy的的实现采取hypergeom_sf
使用对the wikipedia page on the hypergeometric distribution给出的CDF式它。不一定是生存功能的最佳实现)
例如:
In [107]: import mpmath
In [108]: mpmath.mp.dps = 40
In [109]: k, M, n, N
Out[109]: (5007, 45956, 18969, 5267)
In [110]: hypergeom_pmf(k, M, n, N)
Out[110]: mpf('3.897413335837289136238051958307757561884655e-1717')
In [111]: hypergeom_sf(k, M, n, N)
Out[111]: mpf('1.086314878026431217760059547783856962636701e-1718')
https://docs.python.org/2/tutorial/floatingpoint.html德在这里查看浮点数的问题和局限性。 –
在相关说明中,您是通过蟒蛇漂浮物还是numpy漂浮物? –
用于'k','M'和'N'的典型值是什么? –