distributions: support a zero max value in Zipf.

There is no documentation that says zero isn't okay, and the closed interval
[0, k] described by the documentation is perfectly well-defined even when k is
zero. As far as I can tell, there is no reason *not* to support zero: a random
variable that always returns the same value is still a random variable.
absl::Uniform will happily generate on the interval [0, 1) for the same
reason.
PiperOrigin-RevId: 694649518
Change-Id: Ib940406f762a30e27c19c846c45bd908ae8411c3

absl/random/distributions_test.cc

@@ -470,6 +470,13 @@ TEST_F(RandomDistributionsTest, Zipf) {
   EXPECT_NEAR(6.5944, moments.mean, 2000) << moments;
 }
 
+TEST_F(RandomDistributionsTest, ZipfWithZeroMax) {
+  absl::InsecureBitGen gen;
+  for (int i = 0; i < 100; ++i) {
+    EXPECT_EQ(0, absl::Zipf(gen, 0));
+  }
+}
+
 TEST_F(RandomDistributionsTest, Gaussian) {
   std::vector<double> values(kSize);
 

absl/random/zipf_distribution.h

@@ -57,7 +57,7 @@ class zipf_distribution {
    public:
     using distribution_type = zipf_distribution;
 
-    // Preconditions: k > 0, v > 0, q > 1
+    // Preconditions: k >= 0, v > 0, q > 1
     // The precondidtions are validated when NDEBUG is not defined via
     // a pair of assert() directives.
     // If NDEBUG is defined and either or both of these parameters take invalid
@@ -152,7 +152,7 @@ zipf_distribution<IntType>::param_type::param_type(
     : k_(k), q_(q), v_(v), one_minus_q_(1 - q) {
   assert(q > 1);
   assert(v > 0);
-  assert(k > 0);
+  assert(k >= 0);
   one_minus_q_inv_ = 1 / one_minus_q_;
 
   // Setup for the ZRI algorithm (pg 17 of the paper).