Pseudo random number generatio

Linear feedback shift registers are introduced along with the polynomials that completely describe them. The application note describes how they can be implemented and techniques that can be used to improve the statistical properties of the numbers generated.

Introduction

LFSRs (linear feedback shift registers) provide a simple means for generating nonsequential lists of numbers quickly on microcontrollers. Generating the pseudo-random numbers only requires a right-shift operation and an XOR operation. Figure 1 shows a 5-bit LFSR. Figure 2 shows an LFSR implementation in C, and Figure 3 shows a 16-bit LFSR implementation in 8051 assembly.

LFSRs and polynomials

A LFSR is specified entirely by its polynomial. For example, a 6^th-degree polynomial with every term present is represented with the equation x⁶ + x⁵ + x⁴ + x³ + x² + x + 1. There are 2^{(6 - 1)} = 32 different possible polynomials of this size. Just as with numbers, some polynomials are prime or primitive. We are interested in primitive polynomials because they will give us maximum length periods when shifting. A maximum length polynomial of degree n will have 2ⁿ - 1 different states. A new state is transitioned to after each shift. Consequently, a 6^th-degree polynomial will have 31 different states. Every number between 1 and 31 will occur in the shift register before it repeats. In the case of primitive 6^th-degree polynomials, there are only six. Table 1 lists all the primitive 6^th-degree polynomials and their respective polynomial masks. The polynomial mask is created by taking the binary representation of the polynomial and truncating the right-most bit. The mask is used in the code that implements the polynomial. It takes n bits to implement the polynomial mask for an n^th-degree polynomial.

Every primitive polynomial will have an odd number of terms, which means that every mask for a primitive polynomial will have an even number of 1 bit. Every primitive polynomial also defines a second primitive polynomial, its dual. The dual can be found by subtracting the exponent from the degree of the polynomial for each term. For example, given the 6^th-degree polynomial, x⁶ + x + 1, its dual is x^6-6 + x^6-1 + x^6-0, which is equal to x⁶ + x⁵ + 1. In Table 1, polynomials 1 and 2, 3 and 4, 5 and 6 are the duals of each other.

Table 2 lists the period of each different size polynomial and the number of primitive polynomials that exist for each size. Table 3 lists one polynomial mask for each polynomial of a different size. It also shows the first four values that the LFSR will hold after consecutive shifts when the LFSR is initialized to one. This table should help to ensure that the implementation is correct.

The structure of a LFSR

LFSRs can never have the value of zero, since every shift of a zeroed LFSR will leave it as zero. The LFSR must be initialized, i.e., seeded, to a nonzero value. When the LFSR holds 1 and is shifted once, its value will always be the value of the polynomial mask. When the register is all zeros except the most significant bit, then the next several shifts will show the high bit shift to the low bit with zero fill. For example, any 8-bit shift register with a primitive polynomial will eventually generate the sequence 0x80, 0x40, 0x20, 0x10, 8, 4, 2, 1 and then the polynomial mask.

Generating pseudo-random numbers with LFSR

In general, a basic LFSR does not produce very good random numbers. A better sequence of numbers can be improved by picking a larger LFSR and using the lower bits for the random number. For example, if you have a 10-bit LFSR and want an 8-bit number, you can take the bottom 8 bits of the register for your number. With this method you will see each 8-bit number four times and zero, three times, before the LFSR finishes one period and repeats. This solves the problem of getting zeros, but still the numbers do not exhibit very good statistical properties. Instead you can use a subset of the LFSR for a random number to increase the permutations of the numbers and improve the random properties of the LFSR output.

Shifting the LFSR more than once before getting a random number also improves its statistical properties. Shifting the LFSR by a factor of its period will reduce the total period length by that factor. Table 2 has the factors of the periods.

The relatively short periods of the LFSRs can be solved by XORing the values of two or more different sized LFSRs together. The new period of these XORed LFSRs will be the LCM (least common multiple) of the periods. For example, the LCM of a primitive 4-bit and a primitive 6-bit LFSR is the LCM(15, 63), which is 315. When joining LFSRs in this way, be sure to use only the minimum number of bits of the LFSRs; it is a better practice to use less than that. With the 4- and 6-bit LFSRs, no more than the bottom 4 bits should be used. In Figure 2, the bottom 16 bits are used from 32- and 31-bit LFSRs. Note that XORing two LFSRs of the same size will not increase the period.

The unpredictability of the LFSRs can be increased by XORing a bit of "entropy" with the feedback term. Some care should be taken when doing this—there is a small chance that the LFSR will go to all zeros with the addition of the entropy bit. The zeroing of the LFSR will correct itself if entropy is added periodically. This method of XORing a bit with the feedback term is how CRCs (cyclic redundancy checks) are calculated.

Polynomials are not created equal. Some polynomials will definitely be better than others. Table 2 lists the number of primitive polynomials available for bit sizes up to 31 bits. Try different polynomials until you find one that meets your needs. The masks given in Table 3 were randomly selected.

All the basic statistical tests used for testing random number generators can be found in Donald Knuths, The Art of Computer Programming, Volume 2, Section 3.3. More extensive testing can be done using NIST's Statistical Test Suite. NIST also has several publications describing random number testing and references to other test software.


Figure 1. Simplified drawing of a LFSR.


Figure 2. C code implementing a LFSR.


Figure 3. 8051 assembly code to implement a 16-bit LFSR with mask 0D295h.

Table 1. All 6^th-degree primitive polynomials

	Irreducible Polynomial	In Binary Form	Binary Mask	Mask
1	x6 + x + 1	1000011b	100001b	0x21
2	x6 + x5 + 1	1100001b	110000b	0x30
3	x6 + x5 + x2 + x + 1	1100111b	110011b	0x33
4	x6 + x5 + x4 + x + 1	1110011b	111001b	0x39
5	x6 + x5 + x3 + x2 + 1	1101101b	110110b	0x36
6	x6 + x4 + x3 + x + 1	1011011b	101101b	0x2D

Table 2. Polynomial information

Degree	Period	Factors of Period	No. of Primitive Polynomials of This Degree
3	7	7	2
4	15	3, 5	2
5	31	31	6
6	63	3, 3, 7	6
7	127	127	18
8	255	3, 5, 17	16
9	511	7, 73	48
10	1,023	3, 11, 31	60
11	2,047	23, 89	176
12	4,095	3, 3, 5, 7, 13	144
13	8,191	8191	630
14	16,383	3, 43, 127	756
15	32,767	7, 31, 151	1,800
16	65,535	3, 5, 17, 257	2,048
17	131,071	131071	7,710
18	262,143	3, 3, 3, 7, 19, 73	7,776
19	524,287	524287	27,594
20	1,048,575	3, 5, 5, 11, 31, 41	24,000
21	2,097,151	7, 7, 127, 337	84,672
22	4,194,303	3, 23, 89, 683	120,032
23	8,388,607	47, 178481	356,960
24	16,777,215	3, 3, 5, 7, 13, 17, 241	276,480
25	33,554,431	31, 601, 1801	1,296,000
26	67,108,863	3, 2731, 8191	1,719,900
27	134,217,727	7, 73, 262657	4,202,496
28	268,435,455	3, 5 29, 43, 113, 127	4,741,632
29	536,870,911	233, 1103, 2089	18,407,808
30	1,073,741,823	3, 3, 7, 11, 31, 151, 331	17,820,000
31	2,147,483,647	2147483647	69,273,666
32	4,294,967,295	3, 5, 17, 257, 65537	Not Available

Table 3. Sample masks and first four values output after initializing LFSR with one

Degree	Typical Mask	First Four Values in LFSR After Consecutive Shifts
3	0x5	0x5	0x7	0x6	0x3
4	0x9	0x9	0xD	0xF	0xE
5	0x1D	0x1D	0x13	0x14	0xA
6	0x36	0x36	0x1B	0x3B	0x2B
7	0x69	0x69	0x5D	0x47	0x4A
8	0xA6	0xA6	0x53	0x8F	0xE1
9	0x17C	0x17C	0xBE	0x5F	0x153
10	0x32D	0x32D	0x2BB	0x270	0x138
11	0x4F2	0x4F2	0x279	0x5CE	0x2E7
12	0xD34	0xD34	0x69A	0x34D	0xC92
13	0x1349	0x1349	0x1AED	0x1E3F	0x1C56
14	0x2532	0x2532	0x1299	0x2C7E	0x163F
15	0x6699	0x6699	0x55D5	0x4C73	0x40A0
16	0xD295	0xD295	0xBBDF	0x8F7A	0x47BD
17	0x12933	0x12933	0x1BDAA	0xDED5	0x14659
18	0x2C93E	0x2C93E	0x1649F	0x27B71	0x3F486
19	0x593CA	0x593CA	0x2C9E5	0x4F738	0x27B9C
20	0xAFF95	0xAFF95	0xF805F	0xD3FBA	0x69FDD
21	0x12B6BC	0x12B6BC	0x95B5E	0x4ADAF	0x10E06B
22	0x2E652E	0x2E652E	0x173297	0x25FC65	0x3C9B1C
23	0x5373D6	0x5373D6	0x29B9EB	0x47AF23	0x70A447
24	0x9CCDAE	0x9CCDAE	0x4E66D7	0xBBFEC5	0xC132CC
25	0x12BA74D	0x12BA74D	0x1BE74EB	0x1F49D38	0xFA4E9C
26	0x36CD5A7	0x36CD5A7	0x2DABF74	0x16D5FBA	0xB6AFDD
27	0x4E5D793	0x4E5D793	0x6973C5A	0x34B9E2D	0x5401885
28	0xF5CDE95	0xF5CDE95	0x8F2B1DF	0xB25867A	0x592C33D
29	0x1A4E6FF2	0x1A4E6FF2	0xD2737F9	0x1CDDF40E	0xE6EFA07
30	0x29D1E9EB	0x29D1E9EB	0x3D391D1E	0x1E9C8E8F	0x269FAEAC
31	0x7A5BC2E3	0x7A5BC2E3	0x47762392	0x23BB11C9	0x6B864A07
32	0xB4BCD35C	0xB4BCD35C	0x5A5E69AE	0x2D2F34D7	0xA22B4937