2.1 The class of hybrid **LDPC** **codes** 53

• Check node update: Consider a check node c **and** a variable node v. Let {v 1 , . . ., v dc−1}

be the set of all variable nodes connected to c, except v. Let G be the Cartesian product

group of the groups of the variable nodes in {v 1 , . . .,v dc−1}. For all a ∈ G(q v )

l (t)

cv (a) = µ cv

∑

(b 1 ,...,b dc−1 )∈G:

L dc−1

i=1 Av i cb i=A vca

d∏

c−1

n=1

r (t)

v nc(b n ) (2.5)

where µ cv is a normalization factor, **and** the ⊕ operator highlights that the addition

is performed over G(q c ), the group of the row corresponding to c, as defined in

Section 2.1.4.

• Stopping criterion: Consider a variable node v. Let {c 1 , . . .,c dv } be the set of all

check nodes connected to v. Equation (2.6) corresponds to the decision rule on

symbols values, at iteration t:

ˆx (t)

v

= arg maxr (0)

a

v (a) d v

∏

n=1

l c (t)

nv (a) . (2.6)

Variable **and** check node updates are performed **iterative**ly until the decoder has

converged to a codeword, or until the maximum number of iterations is reached.

It is possible to have an efficient Belief propagation decoder for hybrid **LDPC** **codes**.

As mentioned in [11][45], the additive group structure possesses a Fourier transform, so

that efficient computation of the convolution can be done in the Fourier domain. One

**decoding** iteration of BP algorithm for hybrid **LDPC** **codes**, in the probability domain

with a flooding schedule, is composed of:

• Step 1 Variable node update in G(q j ) : pointwise product of incoming messages

• Step 2 Message extension G(q j ) → G(q i ) (see definition 6)

• Step 3 Parity-Check update in G(q i ) in the Fourier domain

◦ FFT of size q i

◦ Pointwise product of FFT vectors

◦ IFFT of size q i

• Step 4 Message truncation from G(q i ) → G(q j ) (see definition 7)

Although we do not focus on low-complexity decoders, it is important to note that hybrid

**LDPC** **codes** are compliant with reduced complexity non-binary decoders which have

been presented recently in the literature [46, 47]. In particular, [46] introduces simplified

**decoding** of GF(q) **LDPC** **codes** **and** shows that they can compete with binary **LDPC**

**codes** even in terms of **decoding** complexity.