Reed–Muller code
Reed-Muller code RM(r,m) | |
---|---|
Named after | Irving S. Reed and David E. Muller |
Classification | |
Type | Linear block code |
Block length | |
Message length | |
Rate | |
Distance | |
Alphabet size | |
Notation | -code |
Reed–Muller codes are error-correcting codes that are used in wireless communications applications, particularly in deep-space communication.[1] Moreover, the proposed 5G standard[2] relies on the closely related polar codes[3] for error correction in the control channel. Due to their favorable theoretical and mathematical properties, Reed–Muller codes have also been extensively studied in theoretical computer science.
Reed–Muller codes generalize the Reed–Solomon codes and the Walsh–Hadamard code. Reed–Muller codes are linear block codes that are locally testable, locally decodable, and list decodable. These properties make them particularly useful in the design of probabilistically checkable proofs.
Traditional Reed–Muller codes are binary codes, which means that messages and codewords are binary strings. When r and m are integers with 0 ≤ r ≤ m, the Reed–Muller code with parameters r and m is denoted as RM(r, m). When asked to encode a message consisting of k bits, where holds, the RM(r, m) code produces a codeword consisting of 2m bits.
Reed–Muller codes are named after David E. Muller, who discovered the codes in 1954,[4] and Irving S. Reed, who proposed the first efficient decoding algorithm.[5]
Description using low-degree polynomials
[edit]Reed–Muller codes can be described in several different (but ultimately equivalent) ways. The description that is based on low-degree polynomials is quite elegant and particularly suited for their application as locally testable codes and locally decodable codes.[6]
Encoder
[edit]A block code can have one or more encoding functions that map messages to codewords . The Reed–Muller code RM(r, m) has message length and block length . One way to define an encoding for this code is based on the evaluation of multilinear polynomials with m variables and total degree at most r. Every multilinear polynomial over the finite field with two elements can be written as follows: The are the variables of the polynomial, and the values are the coefficients of the polynomial. Note that there are exactly coefficients. With this in mind, an input message consists of values which are used as these coefficients. In this way, each message gives rise to a unique polynomial in m variables. To construct the codeword , the encoder evaluates the polynomial at all points , where the polynomial is taken with multiplication and addition mod 2 . That is, the encoding function is defined via
The fact that the codeword suffices to uniquely reconstruct follows from Lagrange interpolation, which states that the coefficients of a polynomial are uniquely determined when sufficiently many evaluation points are given. Since and holds for all messages , the function is a linear map. Thus the Reed–Muller code is a linear code.
Example
[edit]For the code RM(2, 4), the parameters are as follows:
Let be the encoding function just defined. To encode the string x = 1 1010 010101 of length 11, the encoder first constructs the polynomial in 4 variables:Then it evaluates this polynomial at all 16 evaluation points (0101 means :
As a result, C(1 1010 010101) = 1101 1110 0001 0010 holds.
Decoder
[edit]As was already mentioned, Lagrange interpolation can be used to efficiently retrieve the message from a codeword. However, a decoder needs to work even if the codeword has been corrupted in a few positions, that is, when the received word is different from any codeword. In this case, a local decoding procedure can help.
The algorithm from Reed is based on the following property: you start from the code word, that is a sequence of evaluation points from an unknown polynomial of of degree at most that you want to find. The sequence may contains any number of errors up to included.
If you consider a monomial of the highest degree in and sum all the evaluation points of the polynomial where all variables in have the values 0 or 1, and all the other variables have value 0, you get the value of the coefficient (0 or 1) of in (There are such points). This is due to the fact that all lower monomial divisors of appears an even number of time in the sum, and only appears once.
To take into account the possibility of errors, you can also remark that you can fix the value of other variables to any value. So instead of doing the sum only once for other variables not in with 0 value, you do it times for each fixed valuations of the other variables. If there is no error, all those sums should be equals to the value of the coefficient searched. The algorithm consists here to take the majority of the answers as the value searched. If the minority is larger than the maximum number of errors possible, the decoding step fails knowing there are too many errors in the input code.
Once a coefficient is computed, if it's 1, update the code to remove the monomial from the input code and continue to next monomial, in reverse order of their degree.
Example
[edit]Let's consider the previous example and start from the code. With we can fix at most 1 error in the code. Consider the input code as 1101 1110 0001 0110 (this is the previous code with one error).
We know the degree of the polynomial is at most , we start by searching for monomial of degree 2.
-
- we start by looking for evaluation points with . In the code this is: 1101 1110 0001 0110. The first sum is 1 (odd number of 1).
- we look for evaluation points with . In the code this is: 1101 1110 0001 0110. The second sum is 1.
- we look for evaluation points with . In the code this is: 1101 1110 0001 0110. The third sum is 1.
- we look for evaluation points with . In the code this is: 1101 1110 0001 0110. The third sum is 0 (even number of 1).
The four sums don't agree (so we know there is an error), but the minority report is not larger than the maximum number of error allowed (1), so we take the majority and the coefficient of is 1.
We remove from the code before continue : code : 1101 1110 0001 0110, valuation of is 0001000100010001, the new code is 1100 1111 0000 0111
-
- 1100 1111 0000 0111. Sum is 0
- 1100 1111 0000 0111. Sum is 0
- 1100 1111 0000 0111. Sum is 1
- 1100 1111 0000 0111. Sum is 0
One error detected, coefficient is 0, no change to current code.
-
- 1100 1111 0000 0111. Sum is 0
- 1100 1111 0000 0111. Sum is 0
- 1100 1111 0000 0111. Sum is 1
- 1100 1111 0000 0111. Sum is 0
One error detected, coefficient is 0, no change to current code.
-
- 1100 1111 0000 0111. Sum is 1
- 1100 1111 0000 0111. Sum is 1
- 1100 1111 0000 0111. Sum is 1
- 1100 1111 0000 0111. Sum is 0
One error detected, coefficient is 1, valuation of is 0000 0011 0000 0011, current code is now 1100 1100 0000 0100.
-
- 1100 1100 0000 0100. Sum is 1
- 1100 1100 0000 0100. Sum is 1
- 1100 1100 0000 0100. Sum is 1
- 1100 1100 0000 0100. Sum is 0
One error detected, coefficient is 1, valuation of is 0000 0000 0011 0011, current code is now 1100 1100 0011 0111.
-
- 1100 1100 0011 0111. Sum is 0
- 1100 1100 0011 0111. Sum is 1
- 1100 1100 0011 0111. Sum is 0
- 1100 1100 0011 0111. Sum is 0
One error detected, coefficient is 0, no change to current code. We know now all coefficient of degree 2 for the polynomial, we can start mononials of degree 1. Notice that for each next degree, there are twice as much sums, and each sums is half smaller.
-
- 1100 1100 0011 0111. Sum is 0
- 1100 1100 0011 0111. Sum is 0
- 1100 1100 0011 0111. Sum is 0
- 1100 1100 0011 0111. Sum is 0
- 1100 1100 0011 0111. Sum is 0
- 1100 1100 0011 0111. Sum is 0
- 1100 1100 0011 0111. Sum is 1
- 1100 1100 0011 0111. Sum is 0
One error detected, coefficient is 0, no change to current code.
-
- 1100 1100 0011 0111. Sum is 1
- 1100 1100 0011 0111. Sum is 1
- 1100 1100 0011 0111. Sum is 1
- 1100 1100 0011 0111. Sum is 1
- 1100 1100 0011 0111. Sum is 1
- 1100 1100 0011 0111. Sum is 1
- 1100 1100 0011 0111. Sum is 1
- 1100 1100 0011 0111. Sum is 0
One error detected, coefficient is 1, valuation of is 0011 0011 0011 0011, current code is now 1111 1111 0000 0100.
Then we'll find 0 for , 1 for and the current code become 1111 1111 1111 1011.
For the degree 0, we have 16 sums of only 1 bit. The minority is still of size 1, and we found and the corresponding initial word 1 1010 010101
Generalization to larger alphabets via low-degree polynomials
[edit]Using low-degree polynomials over a finite field of size , it is possible to extend the definition of Reed–Muller codes to alphabets of size . Let and be positive integers, where should be thought of as larger than . To encode a message of width , the message is again interpreted as an -variate polynomial of total degree at most and with coefficient from . Such a polynomial indeed has coefficients. The Reed–Muller encoding of is the list of all evaluations of over all . Thus the block length is .
Description using a generator matrix
[edit]![]() | This section may be confusing or unclear to readers. In particular, this section uses dense notation that is not explained well for most readers. (March 2011) |
A generator matrix for a Reed–Muller code RM(r, m) of length N = 2m can be constructed as follows. Let us write the set of all m-dimensional binary vectors as:
We define in N-dimensional space the indicator vectors
on subsets by:
together with, also in , the binary operation
referred to as the wedge product (not to be confused with the wedge product defined in exterior algebra). Here, and are points in (N-dimensional binary vectors), and the operation is the usual multiplication in the field .
is an m-dimensional vector space over the field , so it is possible to write
We define in N-dimensional space the following vectors with length and
where 1 ≤ i ≤ m and the Hi are hyperplanes in (with dimension m − 1):
The generator matrix
[edit]The Reed–Muller RM(r, m) code of order r and length N = 2m is the code generated by v0 and the wedge products of up to r of the vi, 1 ≤ i ≤ m (where by convention a wedge product of fewer than one vector is the identity for the operation). In other words, we can build a generator matrix for the RM(r, m) code, using vectors and their wedge product permutations up to r at a time , as the rows of the generator matrix, where 1 ≤ ik ≤ m.
Example 1
[edit]Let m = 3. Then N = 8, and
and
The RM(1,3) code is generated by the set
or more explicitly by the rows of the matrix:
Example 2
[edit]The RM(2,3) code is generated by the set:
or more explicitly by the rows of the matrix:
Properties
[edit]The following properties hold:
- The set of all possible wedge products of up to m of the vi form a basis for .
- The RM (r, m) code has rank
- RM (r, m) = RM (r, m − 1) | RM (r − 1, m − 1) where '|' denotes the bar product of two codes.
- RM (r, m) has minimum Hamming weight 2m − r.
Proof
[edit]- There are
such vectors and have dimension N so it is sufficient to check that the N vectors span; equivalently it is sufficient to check that .
Let x be a binary vector of length m, an element of X. Let (x)i denote the ith element of x. Define
where 1 ≤ i ≤m.
Then
Expansion via the distributivity of the wedge product gives . Then since the vectors span we have . - By 1, all such wedge products must be linearly independent, so the rank of RM(r, m) must simply be the number of such vectors.
- Omitted.
- By induction.
- The RM(0, m) code is the repetition code of length N =2m and weight N = 2m−0 = 2m−r. By 1 and has weight 1 = 20 = 2m−r.
- If 0 < r < m and if
- RM(r,m − 1) has weight 2m−1−r
- RM(r − 1,m − 1) has weight 2m−1−(r−1) = 2m−r
- then the bar product has weight
Decoding RM codes
[edit]RM(r, m) codes can be decoded using majority logic decoding. The basic idea of majority logic decoding is to build several checksums for each received code word element. Since each of the different checksums must all have the same value (i.e. the value of the message word element weight), we can use a majority logic decoding to decipher the value of the message word element. Once each order of the polynomial is decoded, the received word is modified accordingly by removing the corresponding codewords weighted by the decoded message contributions, up to the present stage. So for a rth order RM code, we have to decode iteratively r+1, times before we arrive at the final received code-word. Also, the values of the message bits are calculated through this scheme; finally we can calculate the codeword by multiplying the message word (just decoded) with the generator matrix.
One clue if the decoding succeeded, is to have an all-zero modified received word, at the end of (r + 1)-stage decoding through the majority logic decoding. This technique was proposed by Irving S. Reed, and is more general when applied to other finite geometry codes.
Description using a recursive construction
[edit]A Reed–Muller code RM(r,m) exists for any integers and . RM(m, m) is defined as the universe () code. RM(−1,m) is defined as the trivial code (). The remaining RM codes may be constructed from these elementary codes using the length-doubling construction
From this construction, RM(r,m) is a binary linear block code (n, k, d) with length n = 2m, dimension and minimum distance for . The dual code to RM(r,m) is RM(m-r-1,m). This shows that repetition and SPC codes are duals, biorthogonal and extended Hamming codes are duals and that codes with k = n/2 are self-dual.
Special cases of Reed–Muller codes
[edit]Table of all RM(r,m) codes for m≤5
[edit]All RM(r, m) codes with and alphabet size 2 are displayed here, annotated with the standard [n,k,d] coding theory notation for block codes. The code RM(r, m) is a -code, that is, it is a linear code over a binary alphabet, has block length , message length (or dimension) k, and minimum distance .
0 | 1 | 2 | 3 | 4 | 5 | m | |
RM(m,m) (2m, 2m, 1) | universe codes | ||||||
RM(5,5) (32,32,1) | |||||||
RM(4,4) (16,16,1) | RM(m − 1, m) (2m, 2m−1, 2) | SPC codes | |||||
RM(3,3) (8,8,1) | RM(4,5) (32,31,2) | ||||||
RM(2,2) (4,4,1) | RM(3,4) (16,15,2) | RM(m − 2, m) (2m, 2m−m−1, 4) | extended Hamming codes | ||||
RM(1,1) (2,2,1) | RM(2,3) (8,7,2) | RM(3,5) (32,26,4) | |||||
RM(0,0) (1,1,1) | RM(1,2) (4,3,2) | RM(2,4) (16,11,4) | |||||
RM(0,1) (2,1,2) | RM(1,3) (8,4,4) | RM(2,5) (32,16,8) | RM(r, m=2r+1) (22r+1, 22r, 2r+1) | self-dual codes | |||
RM(−1,0) (1,0,) | RM(0,2) (4,1,4) | RM(1,4) (16,5,8) | |||||
RM(−1,1) (2,0,) | RM(0,3) (8,1,8) | RM(1,5) (32,6,16) | |||||
RM(−1,2) (4,0,) | RM(0,4) (16,1,16) | RM(1,m) (2m, m+1, 2m−1) | punctured Hadamard codes | ||||
RM(−1,3) (8,0,) | RM(0,5) (32,1,32) | ||||||
RM(−1,4) (16,0,) | RM(0,m) (2m, 1, 2m) | repetition codes | |||||
RM(−1,5) (32,0,) | |||||||
RM(−1,m) (2m, 0, ∞) | trivial codes |
Properties of RM(r,m) codes for r≤1 or r≥m-1
[edit]- RM(0, m) codes are repetition codes of length N = 2m, rate and minimum distance .
- RM(1, m) codes are parity check codes of length N = 2m, rate and minimum distance .
- RM(m − 1, m) codes are single parity check codes of length N = 2m, rate and minimum distance .
- RM(m − 2, m) codes are the family of extended Hamming codes of length N = 2m with minimum distance .[7]
References
[edit]- ^ Massey, James L. (1992), "Deep-space communications and coding: A marriage made in heaven", Advanced Methods for Satellite and Deep Space Communications, Lecture Notes in Control and Information Sciences, vol. 182, Springer-Verlag, pp. 1–17, CiteSeerX 10.1.1.36.4265, doi:10.1007/bfb0036046, ISBN 978-3540558514pdf
- ^ "3GPP RAN1 meeting #87 final report". 3GPP. Retrieved 31 August 2017.
- ^ Arikan, Erdal (2009). "Channel Polarization: A Method for Constructing Capacity-Achieving Codes for Symmetric Binary-Input Memoryless Channels - IEEE Journals & Magazine". IEEE Transactions on Information Theory. 55 (7): 3051–3073. arXiv:0807.3917. doi:10.1109/TIT.2009.2021379. hdl:11693/11695. S2CID 889822.
- ^ Muller, David E. (1954). "Application of Boolean algebra to switching circuit design and to error detection". Transactions of the I.R.E. Professional Group on Electronic Computers. EC-3 (3): 6–12. doi:10.1109/irepgelc.1954.6499441. ISSN 2168-1740.
- ^ Reed, Irving S. (1954). "A class of multiple-error-correcting codes and the decoding scheme". Transactions of the IRE Professional Group on Information Theory. 4 (4): 38–49. doi:10.1109/tit.1954.1057465. hdl:10338.dmlcz/143797. ISSN 2168-2690.
- ^ Prahladh Harsha et al., Limits of Approximation Algorithms: PCPs and Unique Games (DIMACS Tutorial Lecture Notes), Section 5.2.1.
- ^ Trellis and Turbo Coding, C. Schlegel & L. Perez, Wiley Interscience, 2004, p149.
Further reading
[edit]- Shu Lin; Daniel Costello (2005). Error Control Coding (2 ed.). Pearson. ISBN 978-0-13-017973-9. Chapter 4.
- J.H. van Lint (1992). Introduction to Coding Theory. GTM. Vol. 86 (2 ed.). Springer-Verlag. ISBN 978-3-540-54894-2. Chapter 4.5.
External links
[edit]- MIT OpenCourseWare, 6.451 Principles of Digital Communication II, Lecture Notes section 6.4
- GPL Matlab-implementation of RM-codes
- Source GPL Matlab-implementation of RM-codes
- Weiss, E. (September 1962). "Generalized Reed-Muller codes". Information and Control. 5 (3): 213–222. doi:10.1016/s0019-9958(62)90555-7. ISSN 0019-9958.