[简短版:答案是 20。根据寻找 2^e / 10^d 形式的数的良好有理逼近来重新解决问题;然后使用连分数为每个合适的d 和e 找到最佳近似值。]
答案似乎是 20:也就是说,有一些 IEEE 754 binary64 浮点数的示例,其十进制扩展具有 20 连续零,但没有在其十进制扩展中具有 21 连续零(不包括前导和尾随零)。对于 9 字符串也是如此。
对于第一部分,我需要做的就是展示这样一个浮动。 0x1.9527560bfbed8p-1000 的值完全可以表示为二进制 64 浮点数,其十进制扩展包含 20 个零的字符串:
1.4770123739081015758322326613397693800319378788862225686396638475789157389044026850930817635789180868803699741668118826590044503912865915000931065333265410967343958956370955236330760696646247901278074331738806828003156818618589682432778455224012594723731303304343292224317331720902511661748324604219378419442700000000000000000000740694966568985212687104794747958616712153948337746429554804241586090095019654323133732729258896166004754316995632195371041441104566613036026346868128222593894931067078171989365490315525401375255259854894072456336393577718955037826961967325532389800834191597056333066925969522850884268136311674777047673845172073566950098844307658716553833345849153012040436628485227928616281881622762650607683099224232137203216552734375E-301 P>
关于九的部分,0x1.c23142c9da581p-405的十进制扩展包含一个20个九的字符串:
2.12818792307269553358078502102171540639252016258831784842556110831434197718043638405555406495645619729155240037555858106390933161420388023706431461384056688295540725831155392678607931808851292893574214797681879999999999999999999941026584542575391157788777223962620780080784703190447744595561259568772261019375946489162743091583251953125E-122 P>
为了解释我是如何找到上面的数字的,并且为了证明没有连续 21 个零的例子,我们需要更加努力。一个实数在其十进制扩展中具有长字符串 9 或 0,对于某些整数 a 和 d 和实数 eps 的形式为 (a + eps)*10^d,其中 a 非零(我们不妨假设 @ 987654335@ 正)和eps 非零和小。例如,如果0 < abs(eps) < 10^-10 则a + eps 小数点后至少有10 个零(如果eps 为正数),或小数点后有10 个九(如果eps 为负数);乘以10^d 可以移动零或九字符串的位置。
但我们对上述形式的数字感兴趣,它们可以同时表示为 IEEE 754 binary64 浮点数;换句话说,对于整数 b 和 e 满足 2^52 <= b <= 2^53 的整数也是 b*2^e 形式的数字,e 的范围有限(并且一旦我们进入低于正常范围,但我们可以稍后再担心)。
因此,结合这一点,我们正在寻找整数 a、b、d 和 e 的 (a + eps) * 10^d = b * 2^e 的解决方案,使得 eps 很小,a 是正数,@987654355 @(我们稍后会担心 d 和 e 的范围)。重新排列,我们得到eps / b = 2^e / 10^d - a / b。换句话说,我们正在寻找对2^e / 10^d 具有有限分母的良好有理近似值。这是连分数的经典应用:给定d 和e,可以有效地找到分母以2^53 为界的最佳有理逼近。
所以一般的解决策略是:
for each appropriate d and e:
find the best rational approximation a / b to 2^e / 10^d with denominator <= 2^53
if (the error in this rational approximation is small enough):
# we've got a candidate
examine the decimal expansion of b*2^e
我们只有大约 2000 个值要检查 e,最坏的情况是每个这样的 e 有几百个 d,所以整个事情在计算上是非常可行的。
现在详细说明:“足够小”是什么意思?哪个d 和e 是“合适的”?
关于“足够小”:假设我们正在寻找至少有 19 个零或九的字符串,所以我们正在寻找0 < abs(eps) <= 10^-19 的解决方案。因此,对于每个d 和e,找到所有a 和b 这样abs(2^e / 10^d - a / b) <= 10^-19 * 2^-52 就足够了。请注意,由于b 的限制,a / b 只能有一个这样的分数;如果还有另一个这样的a' / b',那么我们就有1 / 2^106 <= 1 / (b *b') <= abs(a / b - a' / b') <= 2 * 10^-19 * 2^-52,这是一个矛盾。因此,如果存在这样的分数,它必然是给定分母界的最佳有理近似值。
对于d 和e:为了覆盖包括次正规的二进制64 范围,我们希望e 的范围从-1126 到971 包括在内。如果d 太大,那么2^e / 10^d 会比2^-53 小很多,没有解决的希望; d <= 16 + floor(e*log10(2)) 是一个实际的界限。如果d 太小(或太负),那么2^e / 10^d 将是一个整数并且没有解决方案;为了避免这种情况,我们需要d > min(e, 0)。
介绍完所有内容后,让我们编写一些代码。 Python 解决方案非常简单,部分归功于 Fraction.limit_deminator 方法的存在,该方法完全可以在限制内找到最佳有理近似值。
from fractions import Fraction
from itertools import groupby
from math import floor, log10
def longest_run(s, c):
"""Length of the longest run of a given character c in the string s."""
runs = [list(g) for v, g in groupby(s, lambda k: k == c) if v]
return max(len(run) for run in runs) if runs else 0
def closest_fraction(d, e):
"""Closest rational to 2**e/10**d with denominator at most 2**53."""
f = Fraction(2**max(e-d, 0) * 5**max(-d, 0), 2**max(0, d-e) * 5**max(0, d))
approx = f.limit_denominator(2**53)
return approx.numerator, approx.denominator
seen = set()
emin = -1126
emax = 971
for e in range(emin, emax+1):
dmin = min(e, 0) + 1
dmax = int(floor(e*log10(2))) + 16
for d in range(dmin, dmax+1):
num, den = closest_fraction(d, e)
x = float.fromhex('0x{:x}p{}'.format(den, e))
# Avoid duplicates.
if x in seen:
continue
seen.add(x)
digits = '{:.1000e}'.format(x).split('e')[0].replace('.','').strip('0')
zero_run = longest_run(digits, '0')
if zero_run >= 20:
print "{} has {} zeros in its expansion".format(x.hex(), zero_run)
nine_run = longest_run(digits, '9')
if nine_run >= 20:
print "{} has {} nines in its expansion".format(x.hex(), nine_run)
那里有很大的性能改进空间(不使用 Python 的 fractions 模块将是一个好的开始:-);就目前而言,运行完成需要几分钟。结果如下:
0x1.9527560bfbed8p-1000 has 20 zeros in its expansion
0x1.fa712b8efae8ep-997 has 20 zeros in its expansion
0x1.515476ae79b24p-931 has 20 nines in its expansion
0x1.a5a9945a181edp-928 has 20 nines in its expansion
0x1.86049d3311305p-909 has 20 zeros in its expansion
0x1.69c08f3dd8742p-883 has 20 zeros in its expansion
0x1.1b41d80091820p-861 has 20 zeros in its expansion
0x1.62124e00b5e28p-858 has 20 zeros in its expansion
0x1.ba96e180e35b2p-855 has 20 zeros in its expansion
0x1.31c5be6377c48p-786 has 20 zeros in its expansion
0x1.7e372dfc55b5ap-783 has 20 zeros in its expansion
0x1.7e89dc1c3860ap-555 has 20 nines in its expansion
0x1.7e89dc1c3860ap-554 has 20 nines in its expansion
0x1.7e89dc1c3860ap-553 has 20 nines in its expansion
0x1.7e89dc1c3860ap-552 has 20 nines in its expansion
0x1.30bd91ea994cbp-548 has 20 zeros in its expansion
0x1.4a5f9de9ee064p-468 has 20 nines in its expansion
0x1.9cf785646987dp-465 has 20 nines in its expansion
0x1.c23142c9da581p-408 has 20 nines in its expansion
0x1.c23142c9da581p-407 has 20 nines in its expansion
0x1.c23142c9da581p-406 has 20 nines in its expansion
0x1.c23142c9da581p-405 has 20 nines in its expansion
0x1.ba431f4e34be9p+738 has 20 nines in its expansion