Ziggurat algorithm

File:Ziggurat.gif

The Ziggurat algorithm used to generate sample values with a normal distribution. (Only positive values are shown for simplicity.) The pink dots are initally uniform-distributed random numbers. The desired distribution function is first segmented into equal areas "A". One layer i is selected at random by the uniform source at the left. Then a random value from the top source is multipled by the width of the chosen layer, and the result is x tested to see which region of the slice it falls into with 3 possible outcomes: 1) (left, solid black region) the sample clearly under the curve and is passed directly to output, 2) (right, vertically striped region) the sample value may lie under the curve, and must be tested further. In that case, a random y value within the chosen layer is generated and compared to f(x). If less, the point is under the curve and the value x is output. If not, (the third case), the chosen point x is rejected and the algorithm is restarted from the beginning.

The ziggurat algorithm is an algorithm to generate random numbers from a non-uniform distribution. It belongs to the class of rejection sampling algorithms and can be used for choosing values from a monotone decreasing probability distribution. It can also be applied to symmetric unimodal distributions such as the normal distribution by choosing a point from one half and then randomly choosing a half. It was developed by George Marsaglia and Wai Wan Tang.

Like most such algorithms, it assumes an underlying source of uniformly-distributed random numbers, typically a pseudo-random number generator. In the common case (97.5% of the time for normal and exponential distributions with typical table sizes), the algorithm consists of generating a random floating-point value, a random table index, performing one table lookup, one multiply and one comparison. This is considerably faster than the two more commonly used methods to generate normally distributed random numbers, the Marsaglia polar method or the Box-Muller transform, which require at least a logarithm and a square root.

On the other hand, the ziggurat algorithm is more complex to implement and requires precomputed tables, so it is best used when large numbers of random values are required.

The ziggurat algorithm is so named because it covers the probability distribution with rectangular segments stacked in decreasing size order, resembling a ziggurat.

Theory of operation

The ziggurat algorithm is a rejection sampling algorithm; it randomly generates a point in a distribution slightly larger than the desired distribution, then tests whether the generated point is inside the desired distribution. If not, it tries again. Given a random point underneath a probability distribution curve, its x coordinate is a random number with the desired distribution.

The distribution the ziggurat algorithm chooses from is made up of n equal-area regions; n−1 rectangles that cover the bulk of the desired distribution, on top of a non-rectangular base that includes the tail of the distribution.

Given a monotone decreasing probability distribution function f(x), defined for all x≥0, the base of the ziggurat is defined as all points inside the distribution and below y₀ = f(x₀). This consists of a rectangular region from (0, 0) to (x₀, y₀), and the (typically infinite) tail of the distribution, where x>x₀.

This layer has area A. On top of this, add a rectangular layer of width x₀ and height A/x₀, so it also has area A. The top of this layer intersects the distribution function at a point (x₁, y₁), where y₁ = f(x₁) = y₀ + A/x₀. This layer includes every point in the distibution function between y₀ and y₁, but (unlike the base layer) also includes points such as (x₀, y₁) which are not in the desired distribution.

Further layers are then stacked on top. To use a precomputed table of size n (n=256 is typical), one chooses x₀ such that x_n−1=0, meaning that the top box, box n−1, reaches the distribution's peak at (0, f(0)) exactly.

Layer i extends vertically from y_i−1 to y_i, and can be divided into two regions horizontally: the (generally larger) portion from 0 to x_i which is entirely contained within the desired distribution, and the (small) portion from x_i to x_i−1, which is only partially contained.

Ignoring for a moment the problem of layer 0, and given uniform random variables U₀ and U₁ ∈ [0,1), the ziggurat algorithm can be described as:

Choose a random layer 0 ≤ i < n.
Let x = U₀x_i−1.
If x < x_i, return x.
Let y = y_i−1 + U₁(y_i−y_i−1).
Compute f(x). If y < f(x), return x.
Otherwise, choose new random numbers and go back to step 1.

Step 1 amounts to choosing a low-resolution y coordinate. Step 3 tests if the x coordinate is clearly within the desired distribution function without knowing more about the y coordinate. If it is not, step 4 chooses a high-resolution y coordinate and step 5 does the rejection test.

With closely spaced layers, the algorithm terminates at step 3 a very large fraction of the time. Note that for the top layer n−1, however, this test always fails, because x_n−1 = 0.

Layer 0 can also be divided into a central region and an edge, but the edge is an infinite tail. To use the same algorithm to check if the point is in the central region, generate a fictitious x₋₁ = A/y₀. This will generate points with x < x₀ with the correct frequency, and in the rare case that layer 0 is selected and x ≥ x₀, use a special fallback algorithm to select a point at random from the tail. Because the fallback algorithm is used less than one time in a thousand, speed is not essential.

Thus, the full ziggurat algorithm for one-sided distributions is:

Choose a random layer 0 ≤ i < n.
Let x = U₀x_i−1
If x < x_i, return x.
If i=0, generate a point from the tail using the fallback algorithm.
Let y = y_i−1 + U₁(y_i−y_i−1).
Compute f(x). If y < f(x), return x.
Otherwise, choose new random numbers and go back to step 1.

For a two-sided distribution, of course, the result must be negated 50% of the time. This can often be done conveniently by choosing U₀ ∈ (−1,1) and, in step 3, testing if |x| < x_i.

Fallback algorithms for the tail

Because the ziggurat algorithm only generates most outputs very rapidly, and requires a fallback algorithm for outliers, it is always more complex than a more direct implementation. The fallback algorithm, of course, depends on the distribution.

For an exponential distribution, the tail looks just like the body of the distribution. One way is to fall back to the most elementary algorithm E=−ln(U₁) and let x=x₀−ln(U₁). Another is to call the ziggurat algorithm recursively.

For a normal distribution, Marsaglia suggests a compact algorithm:

Let x = −ln(U₁)/x₀
Let y = −ln(U₂)
If 2y > x², return x+x₀
Otherwise, go back to step 1.

Since x₀ ≈ 3.5 for typical table sizes, the test in step 3 is almost always successful. Note also that −ln(U) is just a simple way to generate an exponentially distributed random number. If you have a ziggurat exponential distribution generator available, you can use it instead.

Optimizations

The algorithm can be performed efficiently with precomputed tables of x_i and y_i = f(x_i), but there are some modifications to make it even faster:

Nothing in the ziggurat algorithm depends on the probability distribution function being normalized (integral under the curve equal to 1); removing normalizing constants can speed up the computation of f(x).
Most uniform random number generators are based on integer random number generators which return an integer in the range [0, 2³²−1]. A table of 2⁻³²x_i lets you use such numbers directly for U₀.
When computing two-sided distributions using a two-sided U₀ as described earlier, the random integer can be interpreted as a signed number in the range [−2³¹, 2³¹−1] and a scale factor of 2⁻³¹ can be used.
Rather than comparing the result of the multiplication in step 3, it is possible to precompute x_i/x_i−1 and compare U₀ with that directly. If U₀ is an integer random number generator, these limits may be premultiplied by 2³² (or 2³¹, as appropriate) so an integer comparison can be used.
With the above two changes, the table of unmodified x_i values is no longer needed and may be deleted.
When generating IEEE 754 single-precision floating point values, which only have a 24-bit mantissa (including the implicit leading 1), the least-significant bits of a 32-bit integer random number are not used. These bits may be used to select the layer number. (See the references below for a detailed discussion of this.)
The first three steps may be put into an inline function, which can call an out-of-line implementation of the less frequently needed steps.

Generating the tables

It is possible to store the entire table precomputed, or just include the values n, y₀, A, and an implementation of f⁻¹(x) in the source code, and compute the remaining values when initializing the random number generator.

As previously described, you can find x_i = f⁻¹(y_i) and y_i+1 = y_i−A/x_i. Repeat n−1 times for the layers of the ziggurat. At the end, you should have y_n−1=f(0). There will, of course, be some round-off error, but it is a useful sanity test to see that it is acceptably small.

When actually filling in the table values, just assume that x_n−1=0 and y_n−1=f(0), and accept the slight difference in layer n−1's area as rounding error.

Finding x₀ and A

Given an initial (guess at) x₀, you need a way to compute the area t of the tail for which x>x₀. For the exponential distribution, this is just −log(x₀), while for the normal distribution, assuming you are using the unnormalized f(x) = e^−x²/2, this is √(π/2)erfc(x/√2). For more awkward distributions, numerical integration may be required.

With this in hand, from x₀, you can find y₀ = f(x₀), the area t in the tail, and the area of the base layer A = x₀y₀ + t.

The compute the series y_i and x_i as above. If y_i > f(0) for any i < n, then the initial estimate x₀ was too low, leading to too large an area A. If y_n−1 < f(0), then the initial estimate x₀ was too high.

Given this, use a root-finding algorithm (such as the bisection method) to find the value x₀ which produces y_n−1 as close to f(0) as possible. Alternatively, look for the value which makes the area of the topmost layer, x_n−2(f(0)−y_n−2), as close to the desired value A as possible. This saves one evaluation of f⁻¹(x) and is actually the condition of greatest interest.

References

George Marsaglia (2000). "The Ziggurat Method for Generating Random Variables" (PDF). Journal of Statistical Software. 5 (8). Retrieved 2007-06-20. {{cite journal}}: Unknown parameter |coauthors= ignored (|author= suggested) (help) Note that this paper numbers the layers from 1 starting at the top, and makes layer 0 at the bottom a special case, while the explanation above numbers layers from 0 at the bottom.
C implementation of the ziggurat method for the normal density function and the exponential density function, that is essentially a copy of the code in the paper.
Jurgen A. Doornik (2005). "An Improved Ziggurat Method to Generate Normal Random Samples" (PDF). Nuffield College, Oxford. Retrieved 2007-06-20. {{cite journal}}: Cite journal requires |journal= (help) Describes the hazards of using the least-significant bits of the integer random number generator to choose the layer number.
Ziggurat algorithm generates normally distributed random numbers describing the ziggurat algorithm introduced in MATLAB version 5.
David B. Thomas (2007). "Gaussian Random Number Generators" (PDF). ACM Computing Surveys. 39 (to appear). Retrieved 2007-06-20. {{cite journal}}: Unknown parameter |coauthors= ignored (|author= suggested) (help) Tests many algorithms for generating normal (a.k.a. Gaussian) random numbers, and concludes that "when maintaining extremely high noise quality is the first priority, and subject to that constraint, speed is also desired, the Ziggurat method will often be the most appropraite choice."