以下是我的回答:

static struct rand_buffer {
  unsigned v, count;
} buf2, buf3;

void push (struct rand_buffer *buf, unsigned n, unsigned v)
{
  buf->v = buf->v * n + v;
  ++buf->count;
}

#define PUSH(n, v)  push (&buf##n, n, v)

int rand16 (void)
{
  int v = buf2.v & 0xf;
  buf2.v >>= 4;
  buf2.count -= 4;
  return v;
}

int rand9 (void)
{
  int v = buf3.v % 9;
  buf3.v /= 9;
  buf3.count -= 2;
  return v;
}

int rand7 (void)
{
  if (buf3.count >= 2) {
    int v = rand9 ();

    if (v < 7)
      return v % 7 + 1;

    PUSH (2, v - 7);
  }

  for (;;) {
    if (buf2.count >= 4) {
      int v = rand16 ();

      if (v < 14) {
        PUSH (2, v / 7);
        return v % 7 + 1;
      }

      PUSH (2, v - 14);
    }

    // Get a number between 0 & 25
    int v = 5 * (rand5 () - 1) + rand5 () - 1;

    if (v < 21) {
      PUSH (3, v / 7);
      return v % 7 + 1;
    }

    v -= 21;
    PUSH (2, v & 1);
    PUSH (2, v >> 1);
  }
}

它比其他的稍微复杂一点，但我相信它最小化了对rand5的调用。与其他解决方案一样，它有小概率会循环很长时间。

我首先想到的是这个。但我不知道它是否均匀分布。 在python中实现

进口随机 def rand5 (): 返回random.randint(1、5) def rand7 (): 返回((rand5() -1) * rand5()) %7)+1

rand7() = (rand5()+rand5()+rand5()+rand5()+rand5()+rand5()+rand5())%7+1

编辑:这并不奏效。误差约为千分之二(假设是完美的rand5)。桶得到:

value   Count  Error%
1       11158  -0.0035
2       11144  -0.0214
3       11144  -0.0214
4       11158  -0.0035
5       11172  +0.0144
6       11177  +0.0208
7       11172  +0.0144

通过转换到的和

n   Error%
10  +/- 1e-3,
12  +/- 1e-4,
14  +/- 1e-5,
16  +/- 1e-6,
...
28  +/- 3e-11

似乎每增加2就增加一个数量级

BTW:上面的误差表不是通过采样产生的，而是通过以下递归关系产生的:

P [x,n]是给定n次调用rand5，输出=x可能发生的次数。

  p[1,1] ... p[5,1] = 1
  p[6,1] ... p[7,1] = 0

  p[1,n] = p[7,n-1] + p[6,n-1] + p[5,n-1] + p[4,n-1] + p[3,n-1]
  p[2,n] = p[1,n-1] + p[7,n-1] + p[6,n-1] + p[5,n-1] + p[4,n-1]
  p[3,n] = p[2,n-1] + p[1,n-1] + p[7,n-1] + p[6,n-1] + p[5,n-1]
  p[4,n] = p[3,n-1] + p[2,n-1] + p[1,n-1] + p[7,n-1] + p[6,n-1]
  p[5,n] = p[4,n-1] + p[3,n-1] + p[2,n-1] + p[1,n-1] + p[7,n-1]
  p[6,n] = p[5,n-1] + p[4,n-1] + p[3,n-1] + p[2,n-1] + p[1,n-1]
  p[7,n] = p[6,n-1] + p[5,n-1] + p[4,n-1] + p[3,n-1] + p[2,n-1]

(我剽窃了亚当·罗森菲尔德的答案，使其运行速度提高了7%左右。)

假设rand5()返回分布相等的{0,1,2,3,4}中的一个，目标是返回分布相等的{0,1,2,3,4,5,6}。

int rand7() {
  i = 5 * rand5() + rand5();
  max = 25;
  //i is uniform among {0 ... max-1}
  while(i < max%7) {
    //i is uniform among {0 ... (max%7 - 1)}
    i *= 5;
    i += rand5(); //i is uniform {0 ... (((max%7)*5) - 1)}
    max %= 7;
    max *= 5; //once again, i is uniform among {0 ... max-1}
  }
  return(i%7);
}

我们在跟踪这个循环在变量max中所能产生的最大值。如果到目前为止的结果在max%7和max-1之间，那么结果将均匀分布在该范围内。如果不是，则使用余数，余数是0到max%7-1之间的随机数，然后再次调用rand()来生成一个新的数字和一个新的max。然后我们重新开始。

编辑:在这个方程中，期望调用rand5()的次数是x:

x =  2     * 21/25
   + 3     *  4/25 * 14/20
   + 4     *  4/25 *  6/20 * 28/30
   + 5     *  4/25 *  6/20 *  2/30 * 7/10
   + 6     *  4/25 *  6/20 *  2/30 * 3/10 * 14/15
   + (6+x) *  4/25 *  6/20 *  2/30 * 3/10 *  1/15
x = about 2.21 calls to rand5()

只需要缩放第一个函数的输出

0) you have a number in range 1-5
1) subtract 1 to make it in range 0-4
2) multiply by (7-1)/(5-1) to make it in range 0-6
3) add 1 to increment the range: Now your result is in between 1-7

这相当于Adam Rosenfield的解决方案，但对一些读者来说可能更清楚一些。它假设rand5()是一个函数，返回1到5范围内的统计随机整数。

int rand7()
{
    int vals[5][5] = {
        { 1, 2, 3, 4, 5 },
        { 6, 7, 1, 2, 3 },
        { 4, 5, 6, 7, 1 },
        { 2, 3, 4, 5, 6 },
        { 7, 0, 0, 0, 0 }
    };

    int result = 0;
    while (result == 0)
    {
        int i = rand5();
        int j = rand5();
        result = vals[i-1][j-1];
    }
    return result;
}

How does it work? Think of it like this: imagine printing out this double-dimension array on paper, tacking it up to a dart board and randomly throwing darts at it. If you hit a non-zero value, it's a statistically random value between 1 and 7, since there are an equal number of non-zero values to choose from. If you hit a zero, just keep throwing the dart until you hit a non-zero. That's what this code is doing: the i and j indexes randomly select a location on the dart board, and if we don't get a good result, we keep throwing darts.

就像亚当说的，在最坏的情况下，它可以一直运行下去，但从统计上看，最坏的情况永远不会发生。：）