SERIAL CONCATENATION OF INTERLEAVED CONVOLUTIONAL CODES FORMING TURBO-LIKE CODES

Abstract

A serial concatenated coder includes an outer coder and an inner coder. The outer coder irregularly repeats bits in a data block according to a degree profile and scrambles the repeated bits. The scrambled and repeated bits are input to an inner coder, which has a rate substantially close to one.

Description

BACKGROUND

Properties of a channel affect the amount of data that can be handled by the channel. The so-called “Shannon limit” defines the theoretical limit of the amount of data that a channel can carry.

Different techniques have been used to increase the data rate that can be handled by a channel. “Near Shannon Limit Error-Correcting Coding and Decoding: Turbo Codes,” by Berrou et al. ICC, pp 1064-1070, (1993), described a new “turbo code” technique that has revolutionized the field of error correcting codes. Turbo codes have sufficient randomness to allow reliable communication over the channel at a high data rate near capacity. However, they still retain sufficient structure to allow practical encoding and decoding algorithms. Still, the technique for encoding and decoding turbo codes can be relatively complex.

A standard turbo coder 100 is shown in FIG. 1. A block of k information bits is input directly to a first coder 102. A k bit interleaver 106 also receives the k bits and interleaves them prior to applying them to a second coder 104. The second coder produces an output that has more bits than its input, that is, it is a coder with rate that is less than 1. The coders 102, 104 are typically recursive convolutional coders.

Three different items are sent over the channel 150: the original k bits, first encoded bits 110, and second encoded bits 112. At the decoding end, two decoders are used: a first constituent decoder 160 and a second constituent decoder 162. Each receives both the original k bits, and one of the encoded portions 110, 112. Each decoder sends likelihood estimates of the decoded bits to the other decoders. The estimates are used to decode the uncoded information bits as corrupted by the noisy channel.

SUMMARY

A coding system according to an embodiment is configured to receive a portion of a signal to be encoded, for example, a data block including a fixed number of bits. The coding system includes an outer coder, which repeats and scrambles bits in the data block. The data block is apportioned into two or more sub-blocks, and bits in different sub-blocks are repeated a different number of times according to a selected degree profile. The outer coder may include a repeater with a variable rate and an interleaver. Alternatively, the outer coder may be a low-density generator matrix (LDGM) coder.

The repeated and scrambled bits are input to an inner coder that has a rate substantially close to one. The inner coder may include one or more accumulators that perform recursive modulo two addition operations on the input bit stream.

The encoded data output from the inner coder may be transmitted on a channel and decoded in linear time at a destination using iterative decoding techniques. The decoding techniques may be based on a Tanner graph representation of the code.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a schematic diagram of a prior “turbo code” system.

FIG. 2 is a schematic diagram of a coder according to an embodiment.

FIG. 3 is a Tanner graph for an irregular repeat and accumulate (IRA) coder.

FIG. 4 is a schematic diagram of an IRA coder according to an embodiment.

FIG. 5A illustrates a message from a variable node to a check node on the Tanner graph of FIG. 3.

FIG. 5B illustrates a message from a check node to a variable node on the Tanner graph of FIG. 3.

FIG. 6 is a schematic diagram of a coder according to an alternate embodiment.

FIG. 7 is a schematic diagram of a coder according to another alternate embodiment.

DETAILED DESCRIPTION

FIG. 2 illustrates a coder 200 according to an embodiment. The coder 200 may include an outer coder 202, an interleaver 204, and inner coder 206. The coder may be used to format blocks of data for transmission, introducing redundancy into the stream of data to protect the data from loss due to transmission errors. The encoded data may then be decoded at a destination in linear time at rates that may approach the channel capacity.

The outer coder 202 receives the uncoded data. The data may be partitioned into blocks of fixed size, say k bits. The outer coder may be an (n,k) binary linear block coder, where n>k. The coder accepts as input a block u of k data bits and produces an output block v of n data bits. The mathematical relationship between u and v is v=T₀u, where T₀ is an n×k matrix, and the rate of the coder is k/n.

The rate of the coder may be irregular, that is, the value of T₀ is not constant, and may differ for sub-blocks of bits in the data block. In an embodiment, the outer coder 202 is a repeater that repeats the k bits in a block a number of times q to produce a block with n bits, where n=qk. Since the repeater has an irregular output, different bits in the block may be repeated a different number of times. For example, a fraction of the bits in the block may be repeated two times, a fraction of bits may be repeated three times, and the remainder of bits may be repeated four times. These fractions define a degree sequence, or degree profile, of the code.

The inner coder 206 may be a linear rate-1 coder, which means that the n-bit output block x can be written as x=T_Iw, where T_I is a nonsingular n×n matrix. The inner coder 210 can have a rate that is close to 1, e.g., within 50%, more preferably 10% and perhaps even more preferably within 1% of 1.

In an embodiment, the inner coder 206 is an accumulator, which produces outputs that are the modulo two (mod-2) partial sums of its inputs. The accumulator may be a truncated rate-1 recursive convolutional coder with the transfer function 1/(1+D). Such an accumulator may be considered a block coder whose input block [x₁, . . . , x_n] and output block [y₁, . . . , y_n] are related by the formula

y₁=x₁

y ₂ =x ₁ ⊕x ₂

y ₃ =x ₁ ⊕x ₂ ⊕x ₃

y _n =x ₁ ⊕x ₂ ⊕x ₃ ⊕ . . . ⊕x _n

where “⊕” denotes mod-2, or exclusive-OR (XOR), addition. An advantage of this system is that only mod-2 addition is necessary for the accumulator. The accumulator may be embodied using only XOR gates, which may simplify the design.

The bits output from the outer coder 202 are scrambled before they are input to the inner coder 206. This scrambling may be performed by the interleaver 204, which performs a pseudo-random permutation of an input block v, yielding an output block w having the same length as v.

The serial concatenation of the interleaved irregular repeat code and the accumulate code produces an irregular repeat and accumulate (IRA) code. An IRA code is a linear code, and as such, may be represented as a set of parity checks. The set of parity checks may be represented in a bipartite graph, called the Tanner graph, of the code. FIG. 3 shows a Tanner graph 300 of an IRA code with parameters (f₁, . . . , f_j; a), where f_i≧0, Σ_if_i=1 and “a” is a positive integer. The Tanner graph includes two kinds of nodes: variable nodes (open circles) and check nodes (filled circles). There are k variable nodes 302 on the left, called information nodes. There are r variable nodes 306 on the right, called parity nodes. There are r=(kΣ_iif_i)/a check nodes 304 connected between the information nodes and the parity nodes. Each information node 302 is connected to a number of check nodes 304. The fraction of information nodes connected to exactly i check nodes is f_i. For example, in the Tanner graph 300, each of the f₂ information nodes are connected to two check nodes, corresponding to a repeat of q=2, and each of the f₃ information nodes are connected to three check nodes, corresponding to q=3.

Each check node 304 is connected to exactly “a” information nodes 302. In FIG. 3, a=3. These connections can be made in many ways, as indicated by the arbitrary permutation of the ra edges joining information nodes 302 and check nodes 304 in permutation block 310. These connections correspond to the scrambling performed by the interleaver 204.

In an alternate embodiment, the outer coder 202 may be a low-density generator matrix (LDGM) coder that performs an irregular repeat of the k bits in the block, as shown in FIG. 4. As the name implies, an LDGM code has a sparse (low-density) generator matrix. The IRA code produced by the coder 400 is a serial concatenation of the LDGM code and the accumulator code. The interleaver 204 in FIG. 2 may be excluded due to the randomness already present in the structure of the LDGM code.

If the permutation performed in permutation block 310 is fixed, the Tanner graph represents a binary linear block code with k information bits (u₁, . . . , u_k) and r parity bits (x₁, . . . , x_r), as follows. Each of the information bits is associated with one of the information nodes 302, and each of the parity bits is associated with one of the parity nodes 306. The value of a parity bit is determined uniquely by the condition that the mod-2 sum of the values of the variable nodes connected to each of the check nodes 304 is zero. To see this, set x₀=0. Then if the values of the bits on the ra edges coming out the permutation box are

$x_j=x_{j-1}+\sum _{i-1}^Rv_{\left(j-1\right)R+i}$

(V₁, . . . , v_ra), then we have the recursive formulafor j=1, 2, . . . , r. This is in effect the encoding algorithm.

Two types of IRA codes are represented in FIG. 3, a nonsystematic version and a systematic version. The nonsystematic version is an (r,k) code, in which the codeword corresponding to the information bits (u₁, . . . , u_k) is (x₁, . . . , x_r). The systematic version is a (k+r, k) code, in which the codeword is (u₁, . . . , u_k; x₁, . . . , x₄.

The rate of the nonsystematic code is

$R_{n\mathrm{sys}}=\frac{a}{\sum _i{\mathrm{if}}_i}$

The rate of the systematic code is

$R_{\mathrm{sys}}=\frac{a}{a+\sum _i{\mathrm{if}}_i}$

For example, regular repeat and accumulate (RA) codes can be considered nonsystematic IRA codes with a=1 and exactly one f_i equal to 1, say f_q=1, and the rest zero, in which case R_nsys simplifies to R=1/q.

The IRA code may be represented using an alternate notation. Let λ_i be the fraction of edges between the information nodes 302 and the check nodes 304 that are adjacent to an information node of degree i, and let ρ_i be the fraction of such edges that are adjacent to a check node of degree i+2 (i.e., one that is adjacent to i information nodes). These edge fractions may be used to represent the IRA code rather than the corresponding node fractions. Define λ(x)=Σ_iλ_ix¹⁻¹ and ρ(x)=Σ_iρ_ixⁱ⁻¹ to be

$f_i=\frac{{\lambda }_i/i}{\sum _j{\lambda }_j/j}$

the generating functions of these sequences. The pair (λ, ρ) is called a degree distribution. For L(x)=Σ_if_ix_i,

The rate of the systematic IRA code given by the

$L\left(x\right)={\int }_0^x\lambda \left(t\right)t/{\int }_0^1\lambda \left(t\right)t$ $\mathrm{Rate}={\left(1+\frac{\sum _j{\rho }_j/j}{\sum _j{\lambda }_j/j}\right)}^{-1}$

degree distribution is given by

“Belief propagation” on the Tanner Graph realization may be used to decode IRA codes. Roughly speaking, the belief propagation decoding technique allows the messages passed on an edge to represent posterior densities on the bit associated with the variable node. A probability density on a bit is a pair of non-negative real numbers p(0), p(1) satisfying p(0)+p(1)=1, where p(0) denotes the probability of the bit being 0, p(1) the probability of it being 1. Such a pair can be represented by its log likelihood ratio, m=log(p(0)/p(1)). The outgoing message from a variable node u to a check node v represents information about u, and a message from a check node u to a variable node v represents information about u, as shown in FIGS. 5A and 5B, respectively.

The outgoing message from a node u to a node v depends on the incoming messages from all neighbors w of u except v. If u is a variable message node, this outgoing message is

$m\left(u\to v\right)=\sum _{w\ne v}m\left(w\to u\right)+m_0\left(u\right)$

where m₀(u) is the log-likelihood message associated with u. If u is a check node, the corresponding formula is

$\tanh \frac{m\left(u\to v\right)}{2}=\prod _{w\ne v}\tanh \frac{m\left(w\to u\right)}{2}$

Before decoding, the messages m(w→u) and m(u→v) are initialized to be zero, and m₀(u) is initialized to be the log-likelihood ratio based on the channel received information. If the channel is memoryless, i.e., each channel output only relies on its input, and y is the output of the channel code bit u, then m₀(u)=log(p(u=0|y)/p(u=1|y)). After this initialization, the decoding process may run in a fully parallel and local manner. In each iteration, every variable/check node receives messages from its neighbors, and sends back updated messages. Decoding is terminated after a fixed number of iterations or detecting that all the constraints are satisfied. Upon termination, the decoder outputs a decoded sequence based on the messages

$m\left(u\right)=\sum w_m\left(w\to u\right).$

Thus, on various channels, iterative decoding only differs in the initial messages m₀(u). For example, consider three memoryless channel models: a binary erasure channel (BEC); a binary symmetric channel (BSC); and an additive white Gaussian noise (AGWN) channel.

In the BEC, there are two inputs and three outputs. When 0 is transmitted, the receiver can receive either 0 or an erasure E. An erasure E output means that the receiver does not know how to demodulate the output. Similarly, when 1 is transmitted, the receiver can receive either 1 or E. Thus, for the BEC, yε{0, E, 1}, and

$m_0\left(u\right)=\{\begin{array}{ccc}+\infty & \mathrm{if}& y=0\\ 0& \mathrm{if}& y=E\\ -\infty & \mathrm{if}& y=1\end{array}$

In the BSC, there are two possible inputs (0,1) and two possible outputs (0, 1). The BSC is characterized by a set of conditional probabilities relating all possible outputs to possible inputs. Thus, for the BSC yε{0, 1},

$m_0\left(u\right)=\{\begin{array}{ccc}\log \frac{1-p}{p}& \mathrm{if}& y=0\\ -\log \frac{1-p}{p}& \mathrm{if}& y=1\end{array}$

In the AWGN, the discrete-time input symbols X take their values in a finite alphabet while channel output symbols Y can take any values along the real line. There is assumed to be no distortion or other effects other than the addition of white Gaussian noise. In an AWGN with a Binary Phase Shift Keying (BPSK) signaling which maps 0 to the symbol with amplitude √{square root over (Es)} and 1 to the symbol with amplitude −√{square root over (Es)}, output yεR, then

m ₀(u)=4y√{square root over (E_s)}/N₀

where N₀/2 is the noise power spectral density.

The selection of a degree profile for use in a particular transmission channel is a design parameter, which may be affected by various attributes of the channel. The criteria for selecting a particular degree profile may include, for example, the type of channel and the data rate on the channel. For example, Table 1 shows degree profiles that have been found to produce good results for an AWGN channel model.

	TABLE 1
	a	2	3	4
λ2	0.139025	0.078194	0.054485
λ3	0.2221555	0.128085	0.104315
λ5		0.160813
λ6	0.638820	0.036178	0.126755
λ10			0.229816
λ11			0.016484
λ12		0.108828
λ13		0.487902
λ14
λ16
λ27			0.450302
λ28			0.017842
Rate	0.333364	0.333223	0.333218
σGA	1.1840	1.2415	1.2615
σ*	1.1981	1.2607	1.2780
(Eb/N0) * (dB)	0.190	−0.250	−0.371
S.L. (dB)	−0.4953	−0.4958	−0.4958

Table 1 shows degree profiles yielding codes of rate approximately 1/3 for the AWGN channel and with a=2, 3, 4. For each sequence, the Gaussian approximation noise threshold, the actual sum-product decoding threshold and the corresponding energy per bit (E_b)-noise power (N₀) ratio in dB are given. Also listed is the Shannon limit (S.L.).

As the parameter “a” is increased, the performance improves. For example, for a=4, the best code found has an iterative decoding threshold of E_b/N₀=−0.371 dB, which is only 0.12 dB above the Shannon limit.

The accumulator component of the coder may be replaced by a “double accumulator” 600 as shown in FIG. 6. The double accumulator can be viewed as a truncated rate 1 convolutional coder with transfer function 1/(1+D+D²).

Alternatively, a pair of accumulators may be the added, as shown in FIG. 7. There are three component codes: the “outer” code 700, the “middle” code 702, and the “inner” code 704. The outer code is an irregular repetition code, and the middle and inner codes are both accumulators.

IRA codes may be implemented in a variety of channels, including memoryless channels, such as the BEC, BSC, and AWGN, as well as channels having non-binary input, non-symmetric and fading channels, and/or channels with memory.

A number of embodiments have been described. Nevertheless, it will be understood that various modifications may be made without departing from the spirit and scope of the invention. Accordingly, other embodiments are within the scope of the following claims.

Claims

1. A method of encoding a signal, comprising: receiving a block of data in the signal to be encoded, the block of data including information bits;performing a first encoding operation on at least some of the information bits, the first encoding operation being a linear transform operation that generates L transformed bits; andperforming a second encoding operation using the L transformed bits as an input, the second encoding operation including an accumulation operation in which the L transformed bits generated by the first encoding operation are accumulated, said second encoding operation producing at least a portion of a codeword,wherein L is two or more.

2. The method of claim 1, further comprising: outputting the codeword, wherein the codeword comprises parity bits.

3. The method of claim 2, wherein outputting the codeword comprises: outputting the parity bits; andoutputting at least some of the information bits.

4. The method of claim 3, wherein outputting the codeword comprises: outputting the parity bits following the information bits.

5. The method of claim 2, wherein performing the first encoding operation comprises transforming the at least some of the information bits via a low density generator matrix transformation.

6. The method of claim 5, wherein generating each of the L transformed bits comprises mod-2 or exclusive-OR summing of bits in a subset of the information bits.

7. The method of claim 6, wherein each of the subsets of the information bits includes a same number of the information bits.

8. The method of claim 6, wherein at least two of the information bits appear in three subsets of the information bits.

9. The method of claim 6, wherein a number of subsets in which the information bits appear is irregular.

10. The method of claim 2, wherein performing the second encoding operation comprises using a first of the parity bits in the accumulation operation to produce a second of the parity bits.

11. The method of claim 10, wherein outputting the codeword comprises outputting the second of the parity bits immediately following the first of the parity bits.

12. The method of claim 2, wherein performing the second encoding operation comprises performing one of a mod-2 addition and an exclusive-OR operation.

13. A method of encoding a signal, comprising: receiving a block of data in the signal to be encoded, the block of data including information bits; andperforming an encoding operation using the information bits as an input, the encoding operation including an accumulation of mod-2 or exclusive-OR sums of bits in subsets of the information bits, the encoding operation generating at least a portion of a codeword.

14. The method of claim 13, further comprising: outputting the codeword, wherein the codeword comprises parity bits.

15. The method of claim 14, wherein outputting the codeword comprises: outputting the parity bits; andoutputting at least some of the information bits.

16. The method of claim 15, wherein the parity bits follow the information bits in the codeword.

17. The method of claim 13, wherein each of the subsets of the information bits includes a constant number of the information bits.

18. The method of claim 13, wherein at least two of the information bits appear in three subsets of the information bits.

19. The method of claim 13, wherein performing the encoding operation comprises: mod-2 or exclusive-OR adding a first subset of information bits in the collection to yield a first sum;mod-2 or exclusive-OR adding a second subset of information bits in the collection and the first sum to yield a second sum.

20. The method of claim 13, wherein performing the encoding operation further comprises: performing one of the mod-2 addition and the exclusive-OR summing of the bits in the subsets.

21. The method of claim 13, wherein a number of subsets in which the information bits appear is irregular.

22. A method comprising: receiving a collection of information bits;mod-2 or exclusive-OR adding a first subset of information bits in the collection to yield a first parity bit;mod-2 or exclusive-OR adding a second subset of information bits in the collection and the first parity bit to yield a second parity bit; andoutputting a codeword that includes the first parity bit and the second parity bit.

23. The method of claim 22, wherein: the method further comprises mod-2 or exclusive-OR adding additional subsets of information bits in the collection and parity bits to yield additional parity bits; anda number of subsets in which the information bits in the collection appear is irregular.