58
votes

Normally, a random number generator returns a stream of bits for which the probability to observe a 0 or a 1 in each position is equal (i.e. 50%). Let's call this an unbiased PRNG.

I need to generate a string of pseudo-random bits with the following property: the probability to see a 1 in each position is p (i.e. the probability to see a 0 is 1-p). The parameter p is a real number between 0 and 1; in my problem it happens that it has a resolution of 0.5%, i.e. it can take the values 0%, 0.5%, 1%, 1.5%, ..., 99.5%, 100%.

Note that p is a probability and not an exact fraction. The actual number of bits set to 1 in a stream of n bits must follow the binomial distribution B(n, p).

There is a naive method that can use an unbiased PRNG to generate the value of each bit (pseudocode):

generate_biased_stream(n, p):
  result = []
  for i in 1 to n:
    if random_uniform(0, 1) < p:
      result.append(1)
    else:
      result.append(0)
  return result

Such an implementation is much slower than one generating an unbiased stream, since it calls the random number generator function once per each bit; while an unbiased stream generator calls it once per word size (e.g. it can generate 32 or 64 random bits with a single call).

I want a faster implementation, even it it sacrifices randomness slightly. An idea that comes to mind is to precompute a lookup table: for each of the 200 possible values of p, compute C 8-bit values using the slower algorithm and save them in a table. Then the fast algorithm would just pick one of these at random to generate 8 skewed bits.

A back of the envelope calculation to see how much memory is needed: C should be at least 256 (the number of possible 8-bit values), probably more to avoid sampling effects; let's say 1024. Maybe the number should vary depending on p, but let's keep it simple and say the average is 1024. Since there are 200 values of p => total memory usage is 200 KB. This is not bad, and might fit in the L2 cache (256 KB). I still need to evaluate it to see if there are sampling effects that introduce biases, in which case C will have to be increased.

A deficiency of this solution is that it can generate only 8 bits at once, even that with a lot of work, while an unbiased PRNG can generate 64 at once with just a few arithmetic instructions.

I would like to know if there is a faster method, based on bit operations instead of lookup tables. For example modifying the random number generation code directly to introduce a bias for each bit. This would achieve the same performance as an unbiased PRNG.


Edit March 5

Thank you all for your suggestions, I got a lot of interesting ideas and suggestions. Here are the top ones:

  • Change the problem requirements so that p has a resolution of 1/256 instead of 1/200. This allows using bits more efficiently, and also gives more opportunities for optimization. I think I can make this change.
  • Use arithmetic coding to efficiently consume bits from an unbiased generator. With the above change of resolution this becomes much easier.
  • A few people suggested that PRNGs are very fast, thus using arithmetic coding might actually make the code slower due to the introduced overhead. Instead I should always consume the worst-case number of bits and optimize that code. See the benchmarks below.
  • @rici suggested using SIMD. This is a nice idea, which works only if we always consume a fixed number of bits.

Benchmarks (without arithmetic decoding)

Note: as many of you have suggested, I changed the resolution from 1/200 to 1/256.

I wrote several implementations of the naive method that simply takes 8 random unbiased bits and generates 1 biased bit:

  • Without SIMD
  • With SIMD using the Agner Fog's vectorclass library, as suggested by @rici
  • With SIMD using intrinsics

I use two unbiased pseudo-random number generators:

I also measure the speed of the unbiased PRNG for comparison. Here are the results:


RNG: Ranvec1(Mersenne Twister for Graphics Processors + Multiply with Carry)

Method: Unbiased with 1/1 efficiency, SIMD=vectorclass (incorrect, baseline)
Gbps/s: 16.081 16.125 16.093 [Gb/s]
Number of ones: 536,875,204 536,875,204 536,875,204
Theoretical   : 104,857,600

Method: Biased with 1/8 efficiency
Gbps/s: 0.778 0.783 0.812 [Gb/s]
Number of ones: 104,867,269 104,867,269 104,867,269
Theoretical   : 104,857,600

Method: Biased with 1/8 efficiency, SIMD=vectorclass
Gbps/s: 2.176 2.184 2.145 [Gb/s]
Number of ones: 104,859,067 104,859,067 104,859,067
Theoretical   : 104,857,600

Method: Biased with 1/8 efficiency, SIMD=intrinsics
Gbps/s: 2.129 2.151 2.183 [Gb/s]
Number of ones: 104,859,067 104,859,067 104,859,067
Theoretical   : 104,857,600

SIMD increases performance by a factor of 3 compared to the scalar method. It is 8 times slower than the unbiased generator, as expected.

The fastest biased generator achieves 2.1 Gb/s.


RNG: xorshift128plus

Method: Unbiased with 1/1 efficiency (incorrect, baseline)
Gbps/s: 18.300 21.486 21.483 [Gb/s]
Number of ones: 536,867,655 536,867,655 536,867,655
Theoretical   : 104,857,600

Method: Unbiased with 1/1 efficiency, SIMD=vectorclass (incorrect, baseline)
Gbps/s: 22.660 22.661 24.662 [Gb/s]
Number of ones: 536,867,655 536,867,655 536,867,655
Theoretical   : 104,857,600

Method: Biased with 1/8 efficiency
Gbps/s: 1.065 1.102 1.078 [Gb/s]
Number of ones: 104,868,930 104,868,930 104,868,930
Theoretical   : 104,857,600

Method: Biased with 1/8 efficiency, SIMD=vectorclass
Gbps/s: 4.972 4.971 4.970 [Gb/s]
Number of ones: 104,869,407 104,869,407 104,869,407
Theoretical   : 104,857,600

Method: Biased with 1/8 efficiency, SIMD=intrinsics
Gbps/s: 4.955 4.971 4.971 [Gb/s]
Number of ones: 104,869,407 104,869,407 104,869,407
Theoretical   : 104,857,600

For xorshift, SIMD increases performance by a factor of 5 compared to the scalar method. It is 4 times slower than the unbiased generator. Note that this is a scalar implementation of xorshift.

The fastest biased generator achieves 4.9 Gb/s.


RNG: xorshift128plus_avx2

Method: Unbiased with 1/1 efficiency (incorrect, baseline)
Gbps/s: 18.754 21.494 21.878 [Gb/s]
Number of ones: 536,867,655 536,867,655 536,867,655
Theoretical   : 104,857,600

Method: Unbiased with 1/1 efficiency, SIMD=vectorclass (incorrect, baseline)
Gbps/s: 54.126 54.071 54.145 [Gb/s]
Number of ones: 536,874,540 536,880,718 536,891,316
Theoretical   : 104,857,600

Method: Biased with 1/8 efficiency
Gbps/s: 1.093 1.103 1.063 [Gb/s]
Number of ones: 104,868,930 104,868,930 104,868,930
Theoretical   : 104,857,600

Method: Biased with 1/8 efficiency, SIMD=vectorclass
Gbps/s: 19.567 19.578 19.555 [Gb/s]
Number of ones: 104,836,115 104,846,215 104,835,129
Theoretical   : 104,857,600

Method: Biased with 1/8 efficiency, SIMD=intrinsics
Gbps/s: 19.551 19.589 19.557 [Gb/s]
Number of ones: 104,831,396 104,837,429 104,851,100
Theoretical   : 104,857,600

This implementation uses AVX2 to run 4 unbiased xorshift generators in parallel.

The fastest biased generator achieves 19.5 Gb/s.

Benchmarks for arithmetic decoding

Simple tests show that the arithmetic decoding code is the bottleneck, not the PRNG. So I am only benchmarking the most expensive PRNG.


RNG: Ranvec1(Mersenne Twister for Graphics Processors + Multiply with Carry)

Method: Arithmetic decoding (floating point)
Gbps/s: 0.068 0.068 0.069 [Gb/s]
Number of ones: 10,235,580 10,235,580 10,235,580
Theoretical   : 10,240,000

Method: Arithmetic decoding (fixed point)
Gbps/s: 0.263 0.263 0.263 [Gb/s]
Number of ones: 10,239,367 10,239,367 10,239,367
Theoretical   : 10,240,000

Method: Unbiased with 1/1 efficiency (incorrect, baseline)
Gbps/s: 12.687 12.686 12.684 [Gb/s]
Number of ones: 536,875,204 536,875,204 536,875,204
Theoretical   : 104,857,600

Method: Unbiased with 1/1 efficiency, SIMD=vectorclass (incorrect, baseline)
Gbps/s: 14.536 14.536 14.536 [Gb/s]
Number of ones: 536,875,204 536,875,204 536,875,204
Theoretical   : 104,857,600

Method: Biased with 1/8 efficiency
Gbps/s: 0.754 0.754 0.754 [Gb/s]
Number of ones: 104,867,269 104,867,269 104,867,269
Theoretical   : 104,857,600

Method: Biased with 1/8 efficiency, SIMD=vectorclass
Gbps/s: 2.094 2.095 2.094 [Gb/s]
Number of ones: 104,859,067 104,859,067 104,859,067
Theoretical   : 104,857,600

Method: Biased with 1/8 efficiency, SIMD=intrinsics
Gbps/s: 2.094 2.094 2.095 [Gb/s]
Number of ones: 104,859,067 104,859,067 104,859,067
Theoretical   : 104,857,600

The simple fixed point method achieves 0.25 Gb/s, while the naive scalar method is 3x faster, and the naive SIMD method is 8x faster. There might be ways to optimize and/or parallelize the arithmetic decoding method further, but due to its complexity I have decided to stop here and choose the naive SIMD implementation.

Thank you all for the help.

11
Maybe you could read 200*n/8 bytes from random device and then draw vector of samples with skewed distribution from it. This would result in just one call to random device, but n computations of sum of bits.4pie0
You're using arithmetic coding backwards: You are encoding (compressing), not decompressing. So turn your tables around: In the p = 1/4 case, if your first random bit is 1 you can already output biased 00 (if you see a 0 you have to wait), and so on. As @MarkDickinson's benchmark shows, for p = 0.1 you can get more than two output bits per input bit.alexis
I think I swapped the probability of 0 and 1; but hope you see what I mean.alexis
one tiny change would turn that psuedocode into Python.cat
Would any of the close-voters care to elaborate? I don't see how this is too broad or primarily opinion-based.m69 ''snarky and unwelcoming''

11 Answers

29
votes

One thing you can do is to sample from the underlying unbiased generator multiple times, getting several 32-bit or 64-bit words, and then performing bitwise boolean arithmetic. As an example, for 4 words b1,b2,b3,b4, you can get the following distributions:

    expression             | p(bit is 1)
    -----------------------+-------------
    b1 & b2 & b3 & b4      |  6.25%
    b1 & b2 & b3           | 12.50%
    b1 & b2 & (b3 | b4)    | 18.75%
    b1 & b2                | 25.00%
    b1 & (b2 | (b3 & b4))  | 31.25%
    b1 & (b2 | b3)         | 37.50%
    b1 & (b2 | b3 | b4))   | 43.75%
    b1                     | 50.00%

Similar constructions can be made for finer resolutions. It gets a bit tedious and still requires more generator calls, but at least not one per bit. This is similar to a3f's answer, but is probably easier to implement and, I suspect, faster than scanning words for 0xF nybbles.

Note that for your desired 0.5% resolution, you would need 8 unbiased words for one biased word, which would give you a resolution of (0.5^8) = 0.390625%.

25
votes

If you're prepared to approximate p based on 256 possible values, and you have a PRNG which can generate uniform values in which the individual bits are independent of each other, then you can use vectorized comparison to produce multiple biased bits from a single random number.

That's only worth doing if (1) you worry about random number quality and (2) you are likely to need a large number of bits with the same bias. The second requirement seems to be implied by the original question, which criticizes a proposed solution, as follows: "A deficiency of this solution is that it can generate only 8 bits at once, even that with a lot of work, while an unbiased PRNG can generate 64 at once with just a few arithmetic instructions." Here, the implication seems to be that it is useful to generate a large block of biased bits in a single call.

Random-number quality is a difficult subject. It's hard if not impossible to measure, and therefore different people will propose different metrics which emphasize and/or devalue different aspects of "randomness". It is generally possible to trade off speed of random-number generation for lower "quality"; whether this is worth doing depends on your precise application.

The simplest possible tests of random number quality involve the distribution of individual values and the cycle length of the generator. Standard implementations of the C library rand and Posix random functions will typically pass the distribution test, but the cycle lengths are not adequate for long-running applications.

These generators are typically extremely fast, though: the glibc implementation of random requires only a few cycles, while the classic linear congruential generator (LCG) requires a multiply and an addition. (Or, in the case of the glibc implementation, three of the above to generate 31 bits.) If that's sufficient for your quality requirements, then there is little point trying to optimize, particularly if the bias probability changes frequently.

Bear in mind that the cycle length should be a lot longer than the number of samples expected; ideally, it should be greater than the square of that number, so a linear-congruential generator (LCG) with a cycle length of 231 is not appropriate if you expect to generate gigabytes of random data. Even the Gnu trinomial nonlinear additive-feedback generator, whose cycle length is claimed to be approximately 235, shouldn't be used in applications which will require millions of samples.

Another quality issue, which is much harder to test, relates to the independence on consecutive samples. Short cycle lengths completely fail on this metric, because once the repeat starts, the generated random numbers are precisely correlated with historical values. The Gnu trinomial algorithm, although its cycle is longer, has a clear correlation as a result of the fact that the ith random number generated, ri, is always one of the two values ri−3&plus;ri−31 or ri−3&plus;ri−31&plus;1. This can have surprising or at least puzzling consequences, particularly with Bernoulli experiments.

Here's an implementation using Agner Fog's useful vector class library, which abstracts away a lot of the annoying details in SSE intrinsics, and also helpfully comes with a fast vectorized random number generator (found in special.zip inside the vectorclass.zip archive), which lets us generate 256 bits from eight calls to the 256-bit PRNG. You can read Dr. Fog's explanation of why he finds even the Mersenne twister to have quality issues, and his proposed solution; I'm not qualified to comment, really, but it does at least appear to give expected results in the Bernoulli experiments I have tried with it.

#include "vectorclass/vectorclass.h"
#include "vectorclass/ranvec1.h"

class BiasedBits {
  public:
    // Default constructor, seeded with fixed values
    BiasedBits() : BiasedBits(1)  {}
    // Seed with a single seed; other possibilities exist.
    BiasedBits(int seed) : rng(3) { rng.init(seed); }

    // Generate 256 random bits, each with probability `p/256` of being 1.
    Vec8ui random256(unsigned p) {
      if (p >= 256) return Vec8ui{ 0xFFFFFFFF };
      Vec32c output{ 0 };
      Vec32c threshold{ 127 - p };
      for (int i = 0; i < 8; ++i) {
        output += output;
        output -= Vec32c(Vec32c(rng.uniform256()) > threshold);
      }
      return Vec8ui(output);
    }

  private:
    Ranvec1 rng;
};

In my test, that produced and counted 268435456 bits in 260 ms, or one bit per nanosecond. The test machine is an i5, so it doesn't have AVX2; YMMV.

In the actual use case, with 201 possible values for p, the computation of 8-bit threshold values will be annoyingly imprecise. If that imprecision is undesired, you could adapt the above to use 16-bit thresholds, at the cost of generating twice as many random numbers.

Alternatively, you could hand-roll a vectorization based on 10-bit thresholds, which would give you a very good approximation to 0.5% increments, using the standard bit-manipulation hack of doing the vectorized threshold comparison by checking for borrow on every 10th bit of the subtraction of the vector of values and the repeated threshold. Combined with, say, std::mt19937_64, that would give you an average of six bits each 64-bit random number.

17
votes

From an information-theoretic point of view, a biased stream of bits (with p != 0.5) has less information in it than an unbiased stream, so in theory it should take (on average) less than 1 bit of the unbiased input to produce a single bit of the biased output stream. For example, the entropy of a Bernoulli random variable with p = 0.1 is -0.1 * log2(0.1) - 0.9 * log2(0.9) bits, which is around 0.469 bits. That suggests that for the case p = 0.1 we should be able to produce a little over two bits of the output stream per unbiased input bit.

Below, I give two methods for producing the biased bits. Both achieve close to optimal efficiency, in the sense of requiring as few input unbiased bits as possible.

Method 1: arithmetic (de)coding

A practical method is to decode your unbiased input stream using arithmetic (de)coding, as already described in the answer from alexis. For this simple a case, it's not hard to code something up. Here's some unoptimised pseudocode (cough, Python) that does this:

import random

def random_bits():
    """
    Infinite generator generating a stream of random bits,
    with 0 and 1 having equal probability.
    """
    global bit_count  # keep track of how many bits were produced
    while True:
        bit_count += 1
        yield random.choice([0, 1])

def bernoulli(p):
    """
    Infinite generator generating 1-bits with probability p
    and 0-bits with probability 1 - p.
    """
    bits = random_bits()

    low, high = 0.0, 1.0
    while True:
        if high <= p:
            # Generate 1, rescale to map [0, p) to [0, 1)
            yield 1
            low, high = low / p, high / p
        elif low >= p:
            # Generate 0, rescale to map [p, 1) to [0, 1)
            yield 0
            low, high = (low - p) / (1 - p), (high - p) / (1 - p)
        else:
            # Use the next random bit to halve the current interval.
            mid = 0.5 * (low + high)
            if next(bits):
                low = mid
            else:
                high = mid

Here's an example usage:

import itertools
bit_count = 0

# Generate a million deviates.
results = list(itertools.islice(bernoulli(0.1), 10**6))

print("First 50:", ''.join(map(str, results[:50])))
print("Biased bits generated:", len(results))
print("Unbiased bits used:", bit_count)
print("mean:", sum(results) / len(results))

The above gives the following sample output:

First 50: 00000000000001000000000110010000001000000100010000
Biased bits generated: 1000000
Unbiased bits used: 469036
mean: 0.100012

As promised, we've generated 1 million bits of our output biased stream using fewer than five hundred thousand from the source unbiased stream.

For optimisation purposes, when translating this into C / C++ it may make sense to code this up using integer-based fixed-point arithmetic rather than floating-point.

Method 2: integer-based algorithm

Rather than trying to convert the arithmetic decoding method to use integers directly, here's a simpler approach. It's not quite arithmetic decoding any more, but it's not totally unrelated, and it achieves close to the same output-biased-bit / input-unbiased-bit ratio as the floating-point version above. It's organised so that all quantities fit into an unsigned 32-bit integer, so should be easy to translate to C / C++. The code is specialised to the case where p is an exact multiple of 1/200, but this approach would work for any p that can be expressed as a rational number with reasonably small denominator.

def bernoulli_int(p):
    """
    Infinite generator generating 1-bits with probability p
    and 0-bits with probability 1 - p.

    p should be an integer multiple of 1/200.
    """
    bits = random_bits()
    # Assuming that p has a resolution of 0.05, find p / 0.05.
    p_int = int(round(200*p))

    value, high = 0, 1
    while True:
        if high < 2**31:
            high = 2 * high
            value = 2 * value + next(bits)
        else:
            # Throw out everything beyond the last multiple of 200, to
            # avoid introducing a bias.
            discard = high - high % 200
            split = high // 200 * p_int
            if value >= discard:  # rarer than 1 time in 10 million
                value -= discard
                high -= discard
            elif value >= split:
                yield 0
                value -= split
                high = discard - split
            else:
                yield 1
                high = split

The key observation is that every time we reach the beginning of the while loop, value is uniformly distributed amongst all integers in [0, high), and is independent of all previously-output bits. If you care about speed more than perfect correctness, you can get rid of discard and the value >= discard branch: that's just there to ensure that we output 0 and 1 with exactly the right probabilities. Leave that complication out, and you'll just get almost the right probabilities instead. Also, if you make the resolution for p equal to 1/256 rather than 1/200, then the potentially time-consuming division and modulo operations can be replaced with bit operations.

With the same test code as before, but using bernoulli_int in place of bernoulli, I get the following results for p=0.1:

First 50: 00000010000000000100000000000000000000000110000100
Biased bits generated: 1000000
Unbiased bits used: 467997
mean: 0.099675
9
votes

Let's say the probability of a 1 appearing is 6,25% (1/16). There are 16 possible bit patterns for a 4 bit-number: 0000,0001, ..., 1110,1111.

Now, just generate a random number like you used to and replace every 1111 at a nibble-boundary with a 1 and turn everything else to a 0.

Adjust accordingly for other probabilities.

9
votes

You'll get theoretically optimal behavior, i.e. make truly minimal use of the random number generator and be able to model any probability p exactly, if you approach this using arithmetic coding.

Arithmetic coding is a form of data compression that represents the message as a sub-interval of a number range. It provides theoretically optimal encoding, and can use a fractional number of bits for each input symbol.

The idea is this: Imagine that you have a sequence of random bits, which are 1 with probability p. For convenience, I will instead use q for the probability of the bit being zero. (q = 1-p). Arithmetic coding assigns to each bit part of the number range. For the first bit, assign the interval [0, q) if the input is 0, and the interval [q, 1) if the input is 1. Subsequent bits assign proportional sub-intervals of the current range. For example, suppose that q = 1/3 The input 1 0 0 will be encoded like this:

Initially       [0, 1),             range = 1
After 1         [0.333, 1),         range = 0.6666        
After 0         [0.333, 0.5555),    range = 0.2222   
After 0         [0.333, 0.407407),  range = 0.074074

The first digit, 1, selects the top two-thirds (1-q) of the range; the second digit, 0, selects the bottom third of that, and so on. After the first and second step, the interval stradles the midpoint; but after the third step it is entirely below the midpoint, so the first compressed digit can be output: 0. The process continues, and a special EOF symbol is added as a terminator.

What does this have to do with your problem? The compressed output will have random zeros and ones with equal probability. So, to obtain bits with probability p, just pretend that the output of your RNG is the result of arithmetic coding as above, and apply the decoder process to it. That is, read bits as if they subdivide the line interval into smaller and smaller pieces. For example, after we read 01 from the RNG, we will be in the range [0.25, 0.5). Keep reading bits until enough output is "decoded". Since you're mimicking decompressing, you'll get more random bits out than you put in. Because arithmetic coding is theoretically optimal, there's no possible way to turn the RNG output into more biased bits without sacrificing randomness: you're getting the true maximum.

The catch is that you can't do this in a couple of lines of code, and I don't know of a library I can point you to (though there must be some you could use). Still, it's pretty simple. The above article provides code for a general-purpose encoder and decoder, in C. It's pretty straightforward, and it supports multiple input symbols with arbitrary probabilities; in your case a far simpler implementation is possible (as Mark Dickinson's answer now shows), since the probability model is trivial. For extended use, a bit more work would be needed to produce a robust implementation that does not do a lot of floating-point computation for each bit.

Wikipedia also has an interesting discussion of arithmetic encoding considered as change of radix, which is another way to view your task.

8
votes

Uh, pseudo-random number generators are generally quite fast. I'm not sure what language this is (Python, perhaps), but "result.append" (which almost certainly contains memory allocation) is likely slower than "random_uniform" (which just does a little math).

If you want to optimize the performance of this code:

  1. Verify that it is a problem. Optimizations are a bit of work and make the code harder to maintain. Don't do them unless necessary.
  2. Profile it. Run some tests to determine which parts of the code are actually the slowest. Those are the parts you need to speed up.
  3. Make your changes, and verify that they actually are faster. Compilers are pretty smart; often clear code will compile into better code that something complex than might appear faster.

If you are working in a compiled language (even JIT compiled), you take a performance hit for every transfer of control (if, while, function call, etc). Eliminate what you can. Memory allocation is also (usually) quite expensive.

If you are working in an interpreted language, all bets are off. The simplest code is very likely the best. The overhead of the interpreter will dwarf whatever you are doing, so reduce its work as much as possible.

I can only guess where your performance problems are:

  1. Memory allocation. Pre-allocate the array at its full size and fill in the entries later. This ensures that the memory won't need to be reallocated while you're adding the entries.
  2. Branches. You might be able to avoid the "if" by casting the result or something similar. This will depend a lot on the compiler. Check the assembly (or profile) to verify that it does what you want.
  3. Numeric types. Find out the type your random number generator uses natively, and do your arithmetic in that type. For example, if the generator naturally returns 32-bit unsigned integers, scale "p" to that range first, then use it for the comparison.

By the way, if you really want to use the least bits of randomness possible, use "arithmetic coding" to decode your random stream. It won't be fast.

7
votes

One way that would give a precise result is to first randomly generate for a k-bit block the number of 1 bits following the binomial distribution, and then generate a k-bit word with exactly that many bits using one of the methods here. For example the method by mic006 needs only about log k k-bit random numbers, and mine needs only one.

6
votes

If p is close to 0, you can calculate the probability that the n-th bit is the first bit that is 1; then you calculate a random number between 0 and 1 and pick n accordingly. For example if p = 0.005 (0.5%), and the random number is 0.638128, you might calculate (I'm guessing here) n = 321, so you fill with 321 0 bits and one bit set.

If p is close to 1, use 1-p instead of p, and set 1 bits plus one 0 bit.

If p isn't close to 1 or 0, make a table of all 256 sequences of 8 bits, calculate their cumulative probabilities, then get a random number, do a binary search in the array of cumulative probabilities, and you can set 8 bits.

6
votes

Assuming that you have access to a generator of random bits, you can generate a value to compare with p bit by bit, and abort as soon as you can prove that the generated value is less-than or greater-or-equal-to p.

Proceed as follows to create one item in a stream with given probability p:

  1. Start with 0. in binary
  2. Append a random bit; assuming that a 1 has been drawn, you'll get 0.1
  3. If the result (in binary notation) is provably smaller than p, output a 1
  4. If the result is provably larger or equal to p, output a 0
  5. Otherwise (if neither can be ruled out), proceed with step 2.

Let's assume that p in binary notation is 0.1001101...; if this process generates any of 0.0, 0.1000, 0.10010, ..., the value cannot become larger or equal than p anymore; if any of 0.11, 0.101, 0.100111, ... is generated, the value cannot become smaller than p.

To me, it looks like this method uses about two random bits in expectation. Arithmetic coding (as shown in the answer by Mark Dickinson) consumes at most one random bit per biased bit (on average) for fixed p; the cost of modifying p is unclear.

5
votes

What it does

This implementation makes single call to random device kernel module via interface of "/dev/urandom" special character file to get number of random data needed to represent all values in given resolution. Maximum possible resolution is 1/256^2 so that 0.005 can be represented by:

328/256^2,

i.e:

resolution: 256*256

x: 328

with error 0.000004883.

How it does that

The implementation calculates the number of bits bits_per_byte which is number of uniformly distributed bits needed to handle given resolution, i.e. represent all @resolution values. It makes then a single call to randomization device ("/dev/urandom" if URANDOM_DEVICE is defined, otherwise it will use additional noise from device drivers via call to "/dev/random" which may block if there is not enough entropy in bits) to get required number of uniformly distributed bytes and fills in array rnd_bytes of random bytes. Finally it reads number of needed bits per each Bernoulli sample from each bytes_per_byte bytes of rnd_bytes array and compares the integer value of these bits to probability of success in single Bernoulli outcome given by x/resolution. If value hits, i.e. it falls in segment of x/resolution length which we arbitrarily choose to be [0, x/resolution) segment then we note success and insert 1 into resulting array.


Read from random device:

/* if defined use /dev/urandom (will not block),
 * if not defined use /dev/random (may block)*/
#define URANDOM_DEVICE 1

/*
 * @brief   Read @outlen bytes from random device
 *          to array @out.
 */
int
get_random_samples(char *out, size_t outlen)
{
    ssize_t res;
#ifdef URANDOM_DEVICE
    int fd = open("/dev/urandom", O_RDONLY);
    if (fd == -1) return -1;
    res = read(fd, out, outlen);
    if (res < 0) {
        close(fd);
        return -2;
    }
#else
    size_t read_n;
    int fd = open("/dev/random", O_RDONLY);
    if (fd == -1) return -1;
    read_n = 0;
    while (read_n < outlen) {
       res = read(fd, out + read_n, outlen - read_n);
       if (res < 0) {
           close(fd);
           return -3;
       }
       read_n += res;
    }
#endif /* URANDOM_DEVICE */
    close(fd);
    return 0;
}

Fill in vector of Bernoulli samples:

/*
 * @brief   Draw vector of Bernoulli samples.
 * @details @x and @resolution determines probability
 *          of success in Bernoulli distribution
 *          and accuracy of results: p = x/resolution.
 * @param   resolution: number of segments per sample of output array 
 *          as power of 2: max resolution supported is 2^24=16777216
 * @param   x: determines used probability, x = [0, resolution - 1]
 * @param   n: number of samples in result vector
 */
int
get_bernoulli_samples(char *out, uint32_t n, uint32_t resolution, uint32_t x)
{
    int res;
    size_t i, j;
    uint32_t bytes_per_byte, word;
    unsigned char *rnd_bytes;
    uint32_t uniform_byte;
    uint8_t bits_per_byte;

    if (out == NULL || n == 0 || resolution == 0 || x > (resolution - 1))
        return -1;

    bits_per_byte = log_int(resolution);
    bytes_per_byte = bits_per_byte / BITS_PER_BYTE + 
                        (bits_per_byte % BITS_PER_BYTE ? 1 : 0);
    rnd_bytes = malloc(n * bytes_per_byte);
    if (rnd_bytes == NULL)
        return -2;
    res = get_random_samples(rnd_bytes, n * bytes_per_byte);
    if (res < 0)
    {
        free(rnd_bytes);
        return -3;
    }

    i = 0;
    while (i < n)
    {
        /* get Bernoulli sample */
        /* read byte */
        j = 0;
        word = 0;
        while (j < bytes_per_byte)
        {
            word |= (rnd_bytes[i * bytes_per_byte + j] << (BITS_PER_BYTE * j));
            ++j;
        }
        uniform_byte = word & ((1u << bits_per_byte) - 1);
        /* decision */
        if (uniform_byte < x)
            out[i] = 1;
        else
            out[i] = 0;
        ++i;
    }

    free(rnd_bytes);    
    return 0;
}

Usage:

int
main(void)
{
    int res;
    char c[256];

    res = get_bernoulli_samples(c, sizeof(c), 256*256, 328); /* 328/(256^2) = 0.0050 */
    if (res < 0) return -1;

    return 0;
}

Complete code, results.

0
votes

Although this question is 5 years old, I believe I have something of value to add. While SIMD and arithmetic decoding are undoubtedly great techniques, its hard to ignore that the bitwise boolean arithmetic suggested by @mindriot is very simple and easy to grasp.

However, it's not immediately apparent how you would go about efficiently and quickly implementing this solution. For 256 bits (0.00390625) of resolution, you could write a switch statement with 256 cases and then manually determine the required boolean expression by hand for each case. It would take a while to program this but it will compile down to a very fast jump table in C/C++.

But, what if you want 2^16 bits of resolution, or even 2^64? The latter is a resolution of 5.4210109E-20, more precise than most of us would ever need. The task is absolutely impossible by hand, but we can actually construct a small virtual machine to do this quickly in just 30 lines of C code.

Let's construct the machine for 256 bits of resolution. I'll define probability = resolution/256. e.g., when resolution = 64, then probability = 0.25. As it turns out, the numerator (resolution) actually implicitly encodes the required boolean operations in its binary representation.

For example, what expression generates probability = 0.69140625 = 177/256? The resolution is 177, which in binary is 10110001. Let AND = 0 and OR = 1. We start after the first nonzero least significant bit and read toward the most significant bit. Map the 0/1 to AND/OR. Thus, starting from b1 and reading right to left, we generate the boolean expression (((((((b1 and b2) and b3) and b4) or b5) or b6) and b7) or b8). A computer-generated truth table will confirm 177 cases yield True. To give another example, probability = 0.4375 = 112/256 gives the resolution in binary as 01110000. Reading the 3 bits in order after the first non zero LSB (011) gives ((b1 | b2) | b3) & b4).

Since all we need are the two AND and OR operations, and since the resolution encodes the exact boolean expression we need, a virtual machine can be programmed which interprets the resolution as bitcode. AND and OR are just opcodes that act immediately on the output of an unbiased random number generator. Here is my sample C code:

uint64_t rng_bias (uint64_t *state, const uint8_t resolution)
{
    if (state == NULL) return 0;
    
    //registers
    uint64_t R0 = 0;
    uint8_t  PC = __builtin_ctz(resolution|0x80);
    
    //opcodes
    enum
    {
        OP_ANDI = 0,
        OP_ORI  = 1,
    };
    
    //execute instructions in sequence from LSB -> MSB
    while (PC != (uint8_t) 0x8)
    {
        switch((resolution >> PC++) & (uint8_t) 0x1)
        {
            case OP_ANDI:
                R0 &= rng_generator(state);
                break;
                
            case OP_ORI:
                R0 |= rng_generator(state);
                break;
        }
    }
    
    return R0;
}

The virtual machine is nothing more than 2 registers and 2 opcodes. I am using GCC's builtin function ctz which counts the trailing zero bits so that I can easily find the first nonzero LSB. I bitwise-or the ctz argument with 0x80 because passing zero is undefined. Any other decent compiler should have a similar function. Notice, that unlike the examples I showed by hand, the VM interprets the bitcode starting on the first nonzero LSB, not after. This is because I need to make at least one call to the PRNG to generate the base p=0.5 and p=0.0 cases.

The state pointer and rng_generator() calls are used to interface with your random number generator. For example, for demonstration purposes I can use Marsaglia's Xorshift64:

uint64_t rng_generator(uint64_t *state)
{
    uint64_t x = *state;
    
    x ^= x << 13;
    x ^= x >> 7;
    x ^= x << 17;
    
    return *state = x;
}

All the user/you need to do is manage a separate uint64_t state variable, which must be appropriately seeded prior to using either function.

It is extremely easy to scale to 2^64 bits or whatever other arbitrary resolution desired. use ctzll instead for unsigned long long arguments, change the uint8_t types to uint64_t, and change the while loop check to 64 instead of 8. That's it! Now with at most 64 calls to the PRNG, which is fairly fast, we have access to 5.4210109E-20 resolution.

The key here is that we get the bitcode practically for free. No lexing, parsing, or any other typical VM interpreter tasks. The user provides it via the resolution, without ever realizing. As far as they're concerned, its just the numerator of the probability. As far as we, the implementers are concerned, its nothing more than a string of bitcode for our VM to interpret.

Explaining why the bitcode works requires a whole different and much longer essay. In probability theory, the problem is to determine the generating event (the set of all sample points) of a given probability. Not unlike the usual inverse CDF problem for generating random numbers from a density function. In a computer science viewpoint, in the 256 bit resolution case, we are traversing a depth-8 binary tree where each node represents a probability. The parent node is p=0.5. Left traversal indicates AND operations, right traversal indicates OR. The traversal and node depth maps directly to the LSB->MSB bit encoding that we discussed several paragraphs before.