Alias method

Internally, the algorithm consults two tables, a probability table U_i and an alias table K_i (for 1 ≤ i ≤ n). To generate a random outcome, a fair dice is rolled to determine an index into the two tables. Based on the probability stored at that index, a biased coin is then flipped, and the outcome of the flip is used to choose between a result of i and K_i.[4]

More concretely, the algorithm operates as follows:

1. Generate a uniform random variate 0 ≤ x < 1.
2. Let i = ⌊nx⌋ + 1 and y = nx + 1 − i. (This makes i uniformly distributed on {1, 2, ..., n} and y uniformly distributed on [0, 1).)
3. If y < U_i, return i. This is the biased coin flip.
4. Otherwise, return K_i.

An alternative formulation of the probability table, proposed by Marsaglia et al.[5] as the "square histogram" method, uses the condition x < V_i in the third step (where V_i = (U_i + i − 1)/n) instead of computing y.

The distribution may be padded with additional probabilities p_i = 0 to increase n to a convenient value, such as a power of two.

To generate the table, first initialize U_i = np_i. While doing this, divide the table entries into three categories:

• The "overfull" group, where U_i > 1,
• The "underfull" group, where U_i < 1 and K_i has not been initialized, and
• The "exactly full" group, where U_i = 1 or K_i has been initialized.

If U_i = 1, the corresponding value K_i will never be consulted and is unimportant, but a value of K_i = i is sensible.

As long as not all table entries are exactly full, repeat the following steps:

1. Arbitrarily choose an overfull entry U_i > 1 and an underfull entry U_j < 1. (If one of these exists, the other must, as well.)
2. Allocate the unused space in entry j to outcome i, by setting K_j = i.
3. Remove the allocated space from entry i by changing U_i = U_i − (1 − U_j) = U_i + U_j − 1.
4. Entry j is now exactly full.
5. Assign entry i to the appropriate category based on the new value of U_i.

Each iteration moves at least one entry to the "exactly full" category (and the last moves two), so the procedure is guaranteed to terminate after at most n−1 iterations. Each iteration can be done in O(1) time, so the table can be set up in O(n) time.

Vose[3]^: 974 points out that floating-point rounding errors may cause the guarantee referred to in step 1 to be violated. If one category empties before the other, the remaining entries may have U_i set to 1 with negligible error. The solution accounting for floating point is sometimes called the Walker-Vose method or the Vose alias method.

The Alias structure is not unique.

As the lookup procedure is slightly faster if y < U_i (because K_i does not need to be consulted), one goal during table generation is to maximize the sum of the U_i. Doing this optimally turns out to be NP hard,[5]^: 6 but a greedy algorithm comes reasonably close: rob from the richest and give to the poorest. That is, at each step choose the largest U_i and the smallest U_j. Because this requires sorting the U_i, it requires O(n log n) time.

Although the alias method is very efficient if generating a uniform deviate is itself fast, there are cases where it is far from optimal in terms of random bit usage. This is because it uses a full-precision random variate x each time, even when only a few random bits are needed.

One case arises when the probabilities are particularly well balanced, so many U_i = 1 and K_i is not needed. Generating y is a waste of time. For example if p₁ = p₂ = 1⁄2, then a 32-bit random variate x could be used to make 32 choices, but the alias method will only generate one.

Another case arises when the probabilities are strongly unbalanced, so many U_i ≈ 0. For example if p₁ = 0.999 and p₂ = 0.001, then the great majority of the time, only a few random bits are required to determine that case 1 applies. In such cases, the table method described by Marsaglia et al.[5]^: 1–4 is more efficient. If we make many choices with the same probability we can on average require much less than one unbiased random bit. Using arithmetic coding techniques arithmetic we can approach the limit given by the binary entropy function.

• Donald Knuth, The Art of Computer Programming, Vol 2: Seminumerical Algorithms, section 3.4.1.

• http://www.keithschwarz.com/darts-dice-coins/ Keith Schwarz: Detailed explanation, numerically stable version of Vose's algorithm, and link to Java implementation
• https://jugit.fz-juelich.de/mlz/ransampl Joachim Wuttke: Implementation as a small C library.
• https://gist.github.com/0b5786e9bfc73e75eb8180b5400cd1f8 Liam Huang's Implementation in C++
• https://github.com/joseftw/jos.weightedresult/blob/develop/src/JOS.WeightedResult/AliasMethodVose.cs C# implementation of Vose's algorithm.
• https://github.com/cdanek/KaimiraWeightedList C# implementation of Vose's algorithm without floating point instability.