584 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 55, NO.

2, FEBRUARY 2009

Golden Space–Time Block-Coded Modulation

Laura Luzzi, Ghaya Rekaya-Ben Othman, Member, IEEE, Jean-Claude Belfiore, Member, IEEE, and
Emanuele Viterbo, Senior Member, IEEE

Abstract—In this paper, block-coded modulation is used to de- The construction of codes from cyclic division algebras
sign a 2 2 2 multiple-input multiple-output (MIMO) space–time was first introduced in [6]. In the multiple-input mul-
code for slow fading channels. The Golden Code is chosen as the tiple-output (MIMO) case, Belfiore et al. [7] designed the
inner code; the scheme is based on a set partitioning of the Golden
Code using two-sided ideals whose norm is a power of two. In this Golden Code , a full-rate, full-rank, and information-lossless
case, a lower bound for the minimum determinant is given by the code satisfying the nonvanishing determinant condition. The
minimum Hamming distance. The description of the ring structure MIMO codes that achieve these properties were called
of the quotients suggests further optimization in order to improve Perfect Codes in [8] and also studied in [9].
the overall distribution of determinants. Simulation results show In this paper, we focus on the slow block-fading channel,
that the proposed schemes achieve a significant gain over the un-
coded Golden Code. where the fading coefficients are assumed to be constant for a
certain number of time blocks .1 Traditionally, the design of
Index Terms—Coding gain, Golden Code, Reed–Solomon codes, space–time codes has focused either on the case of short block
space–time block codes. lengths, where the quasi-static interval is approximately the
same as the number of transmit antennas [6], [11], [8], [9],
[5], [12], or on the case [1], [13], [12]. Here we con-
I. INTRODUCTION sider the case of moderate block lengths, that falls in between
these two categories.
Even though fading hinders transmission with respect to the
HE wide diffusion of wireless communications has led
T to a growing demand for high-capacity, highly reliable
transmission schemes over fading channels. The use of multiple
AWGN case, fast fading is actually beneficial because the trans-
mission paths at different times can be regarded as independent.
transmit and receive antennas can greatly improve performance On the contrary, with slow fading the ergodicity assumption
because it increases the diversity order of the system, defined must be dropped, leading to a performance loss. This loss can
as the number of independent transmit–receive paths. In order be compensated using coded modulation: in a general setting, a
to exploit fully the available diversity, a new class of code de- full-rank space–time block code is used as an inner code to guar-
signs, called Space–Time Block Codes, was developed in [1]. In antee full diversity, and is combined with an outer code which
the coherent, block fading model, where the channel coefficients improves the minimum determinant.
are supposed to be known at the receiver, and remain constant We will take as inner code the Golden Code: we focus on
for a time block, the fundamental criteria for code design are the problem of designing a block code ,
— the rank criterion, stating that the difference of two distinct where each component is a Golden codeword. We will as-
codewords or “space–time blocks” must be a full-rank ma- sume that the quasi-static interval has length . In order to
trix, increase the minimum determinant, one can consider the ideals
— the determinant criterion, stating that its minimum deter- of . In [14] and [13], a set partitioning of the Golden Code is
minant ought to be maximized [1], [2]. described; it is based on a chain of left ideals , such
Codes meeting these two criteria can be constructed using tools that the minimum determinant in is times that of .
from algebraic number theory. In particular, by choosing a Choosing the components independently in , one ob-
subset of a division algebra over a number field as the code, tains a very simple block code. For small sizes of the signal
one ensures that all the nonzero codewords are invertible. If, constellation these subcodes already yield a performance gain
furthermore, this subset is contained in an order of the algebra, with respect to the “uncoded” Golden Code (that is, with re-
the minimum determinant over all nonzero codewords will be spect to choosing independently). However, the gain
bounded from below and will not vanish when the size of the is cancelled out asymptotically by the loss of rate as the size of
constellation grows to infinity [3]–[5]. the signal set grows to infinity, since an energy increase is re-
Manuscript received December 17, 2007; revised: August 18, 2008. Current quired to maintain the same spectral efficiency, or bit-rate per
version published February 11, 2009. The material in this paper was presented channel use. A better performance is achieved when the are
in part at the 2008 IEEE Information Theory Workshop, Porto, Portugal, May
2008. not chosen in an independent fashion. In [13], two encoders are
L. Luzzi, G. R.-B. Othman, and J.-C. Belfiore are with TELECOM ParisTech, combined: a trellis encoder whose output belongs to the quotient
75013 Paris, France (e-mail: [email protected]; rekaya@telecom-
paristech.fr; [email protected]). 1This kind of behavior might be caused by large obstructions between trans-
E. Viterbo is with DEIS–Università Della Calabria, 87036 Rende (CS), Italy mitter and receiver. The model is realistic if T is smaller than the coherence
(e-mail: [email protected]). time of the channel; for most practical applications, it has been estimated [10]
Communicated by B. S. Rajan, Associate Editor for Coding Theory. that the coherence time is greater than 0.01 s, so that T < 100 is a legitimate
Digital Object Identifier 10.1109/TIT.2008.2009846 assumption.

0018-9448/$25.00 © 2009 IEEE

LUZZI et al.: GOLDEN SPACE–TIME BLOCK-CODED MODULATION 585

, and a lattice encoder for (trellis coded modula-

tion).
The global minimum determinant for the block code is

This expression is difficult to handle because its “mixed terms”

are Frobenius norms of products in . The codes described
in [13] are designed to maximize the approximate parameter
and so a priori they might
be suboptimal; we will consider the mixed terms and so obtain
a tighter bound for .
A rough estimate of the coding gain for the block code comes Fig. 1. The general structure of the encoder.
from its minimum “Hamming distance”, that is, the minimum
number of nonzero components. To increase the Hamming
weight, we will take as our outer code an error correcting code The encoder will be based on a partition of the Golden Code
over the quotient of by one of its ideals. of cardinality :
The paper is organized as follows. After an Overview of the
proposed encoder (Section II), we recall the algebraic construc-
tion of the Golden Code and its properties in Section III. In
Section IV, we describe the general setting for Golden block
codes and the coding gain estimates; in Section V, we study the
two-sided ideals of that are suitable for binary partitioning. Each subset is a coset of , where is a
In Sections VI and VII, we introduce the repetition codes and scalar factor.
the Reed–Solomon block codes over and discuss their perfor- The permutation of the bits in Step b) and the labelling of the
mance obtained through simulations. The interested reader can constellation are chosen in such a way that for each of the
find in the Appendix the main definitions and theorems con- vectors, the coded bits select one of the cosets of and the
cerning quaternion algebras that are cited in the paper. uncoded bits specify a point in the coset.
More precisely, we consider a partition of the QAM constel-
lation into subsets of size (if , the subsets
II. OVERVIEW consist of a single point). Each subset is a scaled version of a
The general structure of the encoder is the following (see -QAM constellation. For instance, a 16-QAM constella-
Fig. 1). tion can be partitioned into four scaled 4-QAM constellations,
a) The binary information message is divided into two parts. as shown in Fig. 8.
A first data block of bits, where , is en- Two examples of block-coded modulation will be described
coded into a block of length by a linear binary code in detail, namely, the following.
• A “repetition code” over the cosets of . In this case
with generator matrix . The second, of
the scaling factor is , the block length is 2, ,
length bits, is left uncoded. This block is optional,
, , so that 4-QAM constellations are used,
that is .
with a partitioning into two BPSK constellations.
b) The two binary sequences are “mixed up” and rearranged
• A “Golden Reed–Solomon” scheme, for which the binary
into vectors of bits each. Each of the
code is derived from an error-correcting code
binary vectors is then used to modulate four signals , ,
over the finite field , simply by representing the ele-
, belonging a QAM constellation of size . (We
ments of the finite field as polynomials of degree
only considered the case of square constellations.)
with binary coefficients. In particular, we consider two
c) Finally, each of the quadruples is used to
cases for a scaling factor :
encode a Golden Codeword
— In the first case we use 4-QAM constellations, and the
uncoded part is empty ; the partition
(1) consists of individual points.
— In the second, we consider 16-QAM constellations par-
where , . titioned into four 4-QAM constellations .
The Golden Code is full-rate and transmits two symbols
per channel use; each symbol carries bits. The III. THE GOLDEN CODE
total information rate of the binary code is . The Since we are interested in the partitioning of the Golden Code,
rate of the block code will be bpcu. we begin by recalling its algebraic construction. For the sake of
586 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 55, NO. 2, FEBRUARY 2009

simplicity, definitions and theorem statements are collected in The Golden Code is defined as . Every codeword in
Appendix B. is of the form , with : see (1).
The Golden Code , introduced in [7], is a full rate,
full-diversity, information lossless and DMT achieving Remark 1: We have seen that , .
code for two transmit and two or more receive antennas. Consequently, , .
This code is constructed using the cyclic division algebra In fact, if ,
over the number field , where
is the golden number. The set is the -vector , since .
space , where is such that , The code has cubic shaping: it is isometric to the cubic lat-
. tice (and so it is information lossless). In fact, if we con-
Here we denote by the canonical conjugacy sending an el- sider the linear mapping that vectorizes matrices
ement to , where

then , where is the unitary matrix

As its degree over its center is , is also called a
quaternion algebra. If we choose , is not a norm in
, and this implies that is a division algebra (see (6)
Theorem 7 in Appendix B) [7].
is a splitting field for (see Theorem 8 in
Appendix B), and so is isomorphic to a subalgebra of Even though is defined (up to a scaling constant) as a right
. The inclusion is given by ideal, it is easy to see that actually it is a two-sided ideal: if
,
(2)

That is, every element admits a matrix representation observing that . But

(3) (7)

is an homomorphism of -modules that maps into itself

The Golden Code is a subring of having two additional bijectively, therefore .
properties: the minimum determinant Finally, is an integral ideal because it is contained in .
Remark 2: For the sake of simplicity, in this section we have
described the Golden Code as an infinite code. However in a
practical transmission scheme, one considers a finite subset of
should be strictly bounded away from 0, and moreover the code
, by choosing the information symbols , , , in a QAM
is information lossless.
constellation carved from .
For the first condition, if one requires that the matrix elements
of belong to the ring of integers of , then
belongs to the -order IV. GOLDEN BLOCK CODES

(4) A. System Model

We consider a slow block-fading channel, where the channel
Since implies that the reduced norm coefficients remain constant during the transmission of code-
belongs to , we have , so for words. The transmitted signal will be a
every . vector of Golden codewords in a block code . The re-
Each codeword of carries two symbols , ceived signal is given by
in , or equivalently four information symbols
: the code is full-rate. (8)
In order to have an information lossless code, a right principal
ideal of of the form was used, where : its matrix where the entries of are independent and identically
representation is distributed (i.i.d.) complex Gaussian random variables with zero
mean and variance per real dimension equal to , and is
the complex Gaussian noise with i.i.d. entries of zero mean and
variance . We consider the coherent case, where the channel
(5)
matrix is known at the receiver.
LUZZI et al.: GOLDEN SPACE–TIME BLOCK-CODED MODULATION 587

The pairwise error probability is bounded by [1] Lemma 1:

(9)

In the above formula, is the average energy per symbol of (14)

and
The proof can be found in Appendix A.
We also state some simple properties of the quaternionic con-
jugate and of the Frobenius norm that will be useful in the se-
quel.
In order to minimize the PEP for a given SNR, we should max- Remark 3:
imize . We will show that a) If , .
b) Let , be two complex-valued matrices. Then

where is the number of nonzero codewords in

(a sort of “Hamming weight”), and (15)
is the minimum square determinant of the Golden Code.
If we simply choose independently in the Golden In particular
Code, the code performance will be poor compared to the fast
(16)
block fading model. We call this scheme the “uncoded Golden
Code”: in this case , for any length . To compare c) If ,
the error probability of a block code with that of the uncoded
Golden Code of equal length with the same data rate, we can (17)
employ the asymptotic coding gain defined in [13]
From (15), it follows that the determinant is bounded from
(10) below by the squared Hamming weight.
Lemma 2: Let . Then
where , and , are the minimum determi-
nants and average constellation energies of the block code and
the uncoded case, respectively.
In all the cases that we considered, the theoretical gain
turned out to be smaller than the actual gain evidenced by com- where is the Hamming
puter simulations. This is not surprising, since is only a com- weight of the block .
parison of the dominant terms in the pairwise error probability.
V. TWO-SIDED IDEALS OF
B. Estimates of the Frobenius Norm
The choice of a good block code of length will be based on
First of all, we give a more explicit expression for a partition chain of ideals of the Golden Code. We would like
. to obtain a binary partition, which is simpler to use for coding
We define the quaternionic conjugacy in the algebra and fully compatible with the choice of a QAM constellation:
we must then use ideals whose index is a power of , that is,
whose norm is a power of .
A similar construction appears in [13] and employs one-sided
Observe that , ideals.
We prefer to choose two-sided ideals; this is a necessary and
(11)
sufficient condition for the quotient group to be also a ring (see
(12) [15, Th. 2.7]). Moreover, two-sided ideals are also invariant
(13) under involution (see Theorem 15 in Appendix C and the fol-
lowing Remark). This will be useful for code optimization in
where denotes the identity matrix.
Section VI.
Recall that the Frobenius norm of a matrix is
In this section we describe the structure of the two-sided
ideals of whose norm is a power of . Unfortunately,
it turns out that the only two-sided ideals with this property
are the trivial ones. We then study the corresponding quotient
Then the following formula holds. rings, which are rings of matrices over nonintegral rings.
588 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 55, NO. 2, FEBRUARY 2009

For these constructions, we will need some notions from non- We define an homomorphism of -vector spaces
commutative algebra (see Appendix C), relating the existence of by specifying the images of
two-sided ideals to the ramification of primes over the base field. the basis
As we have seen in Section III, is a
-order of , and is a two-sided principal
ideal of .
is also a prime ideal since is a
prime ideal of (see Theorem 13 in Appendix C).
Observe that the prime ideals and of are
both ramified in : in fact It is one-to-one since , , , are linearly
independent. To prove that is also a ring homomorphism, it is
sufficient to verify that for all pairs of
and where
basis vectors , .
Recall that as a -lattice, is isometric to , and a
(Remark that , ).
canonical basis is given by . The corresponding
It has been shown in [3] and [5], Section V, that is a max-
elements , , , of are
imal order and its reduced discriminant is : consequently,
from Proposition 11 we learn that and are the
only ramified primes in .
Then Theorem 15 implies that the prime two-sided ideals of
are either of the form , where is prime in , or belong (19)
to .
It follows that the only two-sided ideals of whose norm is It is easy to check that the only invertible elements in
a power of are the trivial ideals of the form . are

A. The Quotient Ring

In the sequel, we will denote by the integral ideal . Observe that the lifts to of non-invertible elements have a
Consider the prime ideal . and are coprime higher determinant.
ideals, that is ; as a consequence,
Remark 4: If is noninvertible

Recall the following basic result:

Theorem 3: (Third isomorphism theorem for rings). Proof: is noninvertible in if and only if
Let and be ideals in a ring . Then . its determinant is noninvertible in , that is,
. (If , , since is a
Taking and , we get division ring.)
Then .
(18)
B. The Quotient Ring
If and are the Again, and are coprime and so ,
canonical projections on the quotient, the ring isomorphism in ; from the third isomorphism theorem for rings, .
(18) is simply given by .
Theorem 13 implies that is a simple algebra over Lemma 6: is isomorphic to the ring of
. We denote the image of through matrices over the ring .
with . Proof: First of all, Lemma 5 implies that is a free
-module, that is a free -module, of dimension . As in
Lemma 4: is isomorphic to the ring of the previous case, we can construct an explicit homomorphism
2 2 matrices over . of -modules :
Proof: We use the well-known lemma [15]:
Lemma 5: Let be a ring with identity, a proper ideal of
, a free -module with basis and the
canonical projection. Then is a free -module with
basis and .
We know that is a -module; the lemma implies One can easily check that is bijective (the images of the basis
that it is also a free -module, that is a vector space elements being linearly independent) and that it is a ring homo-
over , whose basis is . morphism.
LUZZI et al.: GOLDEN SPACE–TIME BLOCK-CODED MODULATION 589

To find an explicit isomorphism between and , VI. THE REPETITION CODE

consider the following diagram, where is the
projection on the quotient, is given by the third isomorphism Here we consider the case where , and the linear
theorem for rings, and is the mapping code is simply the repetition code of length over . Recall
defined in Lemma 6 that the quotient group is a ring because is a two-sided
ideal of .
If is the projection on the quotient ring
), we define
The basis of as a -module is also a basis
of as an -module. The isomorphism is simply the
composition of the inclusion and the quotient mod
. We can compute the images through of the basis vectors:
observing that A. The Minimum Determinant

Recall that as we have seen in Lemma 1

we get

(20)
With the code , we have . In fact if (re-
spectively, ) is a codeword of Hamming weight , clearly
(21) and is greater than the
minimum square determinant in , which is . If on the
Also in this case, the lifts of non-invertible elements of contrary ,
in will have non-invertible determinant, that is
.

C. The Encoder because of (15).

By choosing any bijection of the quotient ring
The codes that we consider follow the general outline of in itself, one obtains a simple variation of the repetition scheme
Forney’s coset codes [16], taking advantage of the decompo-
sition , where is or , and
denotes a set of coset leaders.
— An linear code over or operates
on part of the information data, and these coded bits are Remark 5: A suitable choice of can slightly improve perfor-
used to select . mance. In the case of the repetition code, suppose that
— The remaining information bits are left uncoded and used .
to select .
— If is invertible in , then
— The corresponding block codeword is
in the basis (19), and so the minimum de-
, where is the coset leader of .
terminant of a codeword is also 1,
For a coset code, is bounded by the minimum determinant
and the minimum of is . Thus
of and the minimum distance of the linear code
.
— If on the other side corresponds to a non-invertible,
(22) nonzero element in , then (see Remark 4)

In fact, if , then , and for

, . If on the contrary
, there are at least components of
and
which do not belong to , and consequently are nonzero, and
. .
So the performance of a coset code will be always limited by This remark suggests that it might be more convenient to con-
the minimum determinant of , except if the code on is the sider a group homomorphism which
zero code. maps invertible elements into non-invertible elements, raising
If is simply or , the set of possible coordi- the minimum determinant to if invertible, non-in-
nates for the coset leaders of in coincides with vertible: , but (see Remark 3)
the (BPSK) and (4-QAM) constellations respectively. This and so , and .
makes it much easier to implement coset codes with high Ham- Such a function is not difficult to define, and in the case
ming distance. of 4-QAM modulation, an exhaustive search on the finite lat-
590 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 55, NO. 2, FEBRUARY 2009

tice shows that the distribution of determinants for is indeed

better.2

B. The Encoder
Only 4 bits are needed to select an element of
, while the number of bits needed to select an ele-
ment in the ideal depends on the chosen modulation scheme.
Using 4-QAM constellations, the two choices of an element in
require 4 bits each: in total, each codeword carries 12
information bits, yielding a spectral efficiency of 3 bpcu.
Suppose that is the binary input.
— are used to select the matrix
in the basis (19). The corre-
sponding element of is
.
— are used to select two codewords in :
,
. Fig. 2. Performance of the repetition code C and of the variation C at 3
bpcu compared with the uncoded Golden Code scheme with the same spectral
— The final block codeword is . efficiency. The channel is supposed to be constant for two time blocks.

C. Asymptotic Coding Gain

Since the minimum determinant does not change, the asymp- the minimum number of noninvertible components also have a
totic coding gain estimate is the same for all choices of . significant influence on the coding gain of a block code design.
We compare these schemes with the uncoded Golden Code Ideally, in order to keep track of these parameters, one ought
at 3 bpcu, using 4-QAM constellations for the symbols , to employ error-correcting codes on . However, at
and BPSK constellations for the symbols , in each Golden present very little is known about codes over noncommutative
codeword (see (1)). The average energy per symbol is rings; we choose shortened Reed–Solomon codes instead be-
, and cause they are maximum distance separable and their imple-
mentation is very simple; we will restrict our attention to the
additive structure, defining a group isomorphism between
and the finite field
This computation gives a theoretical gain of at least
dB dB. A. The 4-QAM Case

D. Simulation Results Using 4-QAM constellations to modulate each of the four

symbols , , , in a Golden codeword (1), we obtain a total
Fig. 2 shows the performance of the codes and , which of 256 codewords, one in each coset of .
gain 2.4 and 2.9 dB, respectively, over the uncoded scheme at 3 We consider an Reed–Solomon code over .
bpcu at the frame error rate of , supposing that the channel Each quadruple of 4-QAM signals carries 8 bits or
is constant for two time blocks. one byte; each block of Golden codewords will carry bytes,
corresponding to information bytes.
The encoding procedure involves several steps.
VII. GOLDEN REED–SOLOMON CODES a) Reed–Solomon encoding:
The repetition code has the advantage of simplicity, but Each information byte can be seen as a binary polyno-
clearly its performance is limited by the fact that the minimum mial of degree , that is, an element of the Galois Field
Hamming distance is only . To increase the Hamming dis- . An information message of bytes, seen as a vector
tance, we need to use a more sophisticated error-correcting , is encoded into a code-
code. word using the
As we have seen in the previous sections, in addition to the shortened code . For our purposes, it is much better to
minimum Hamming distance, the multiplicative structure and use a systematic version of the code that preserves the first
 (e ) = e + e + e , h
2In fact, if we define h  (e ) = e + e + e , bits of the input.
h (e ) = e + e + e , h (e ) = e + e + e with respect to the basis (19), b) From the Galois field to the matrix ring :
we have We can represent the elements of as bytes,
q = 1 + 66q + 120q + 48q + 202q + 111 simply by vectorizing each matrix and separating real and
imaginary parts. Since we are only working with the addi-
q = 1 + 24q + 61q + 24q + 8q + 74q +111: tive structure, we can identify and , which
are both -vector spaces of dimension . According to
our simulation results, it seems that the choice of the linear
LUZZI et al.: GOLDEN SPACE–TIME BLOCK-CODED MODULATION 591

identification has very little influence on the code perfor- c) Intermediate steps: At each iteration of the algorithm, the
mance. triple currently at the top of
c) From the matrix ring to the quotient ring the stack is examined.
: • If , its “children” nodes
For this step, we make use of the isomorphism of
-modules de-
scribed in Section V-B that relates the coordinates with for
respect to the bases and (20). Let
be the coordinates of a codeword in are generated, updating the corresponding Euclidean
the basis . distances
d) Golden Code encoding:
For each of the vector components, the symbols
correspond to four 4-QAM sig- The “parent” node is deleted from the stack and the
nals, and can be encoded into a Golden codeword of children are inserted in the stack and sorted with respect
the form (1). Thus we have obtained a Golden block to distance, or discarded if the distance is greater than
, where .
is injective. (Knowing the minimum distances component-wise,
B. Decoding one can require a stronger condition without losing
optimality, namely, ).
ML decoding consists in the search for the minimum of the
• If , generate the Reed–Solomon codeword
Euclidean distance
and store in the stack
(recall that is an initial segment of ), where

over all the images of Reed–Solomon codewords.

One can first compute and store in memory the Euclidean
distances • If , the search terminates and the initial segment
(23) of length of is the decoded message.
d) Choice of the cost function : A simple bound for the
for every component of the received vector and decoder may be the distance from the received signal of
for all the Golden codewords , that can be the (unique) Golden-RS codeword corresponding to the
obtained from a quadruple of 4-QAM symbols. “closest choice” for the first
The search for the minimum can be carried out using the components. Any subset of components may be used
Viterbi algorithm or a tree search algorithm. as well to improve the minimum provided that the corre-
Stack Decoding: For our computer simulations, we have sponding lines in the Reed–Solomon generator matrix are
chosen to use a stack decoding algorithm. If the code is based linearly independent.
on an Reed–Solomon code with systematic gener-
ator matrix, the codewords are the possible paths in a C. Simulation Results
full tree with height and 256 outgoing branches per node. In the 4-QAM case, the spectral efficiency of the Golden
The decoder will store in a stack a certain number of triples Reed–Solomon codes is given by
, where is an incomplete path of length in the tree,
and is its distance from the initial segment of . bits
bpcu
An upper bound for the minimum distance of the received channel uses
point to the lattice of Golden-RS codewords will be used as a From Lemma 2, it follows that using an
“cost function” for the stack. Reed–Solomon code, .
a) Sorting of distances: Before the search, for each compo- If , the spectral efficiency is 2 bpcu. Comparing the
nent , the distances of (23) are sorted 4-QAM, Golden-RS design with the
in increasing order: let uncoded Golden Code using BPSK , the asymp-
totic coding gain is
be the resulting sequence.
(24)
b) First step: At the beginning, the initial segments of length
1 are inserted into a previously empty stack: the triples
Figs. 3 and 4 show the performance comparisons of the
Golden-RS codes and with the corresponding
uncoded schemes at the spectral efficiency of 2 bpcu.
are entered in decreasing order with respect to the Assuming the channel to be constant for 4 blocks and 6
distance, discarding those whose distances are greater blocks, respectively, the Golden-RS codes outperform the
than . uncoded scheme by 6.1 dB and 7.0 dB.
592 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 55, NO. 2, FEBRUARY 2009

Fig. 3. Comparison between suboptimal decoding and ML decoding for the Fig. 5. Performance of (4; 2; 3), (8; 4; 5), and (12; 6; 7) Golden
RS(4; 2; 3) code at 2 bpcu. The first method achieves a gain of only 1.1 dB Reed–Solomon codes with suboptimal decoding at 2 bpcu compared to the
over the uncoded case, compared to the 6.1 dB of the second. uncoded Golden Code scheme with the same spectral efficiency.

D. Suboptimal Decoding
One can replace maximum likelihood (ML) decoding with
separate Sphere Decoders on each of the components of
. The signal is then demodulated, and mapped to a vector
in using the inverse mappings of Steps c)
and b) in Section VII-A. The received sequence
does not necessarily belong to the RS code, so a final step of
RS decoding is needed. This “hard” decoding has the advan-
tage of speed and allows to use longer Reed–Solomon codes
with high minimum distance. However it is highly suboptimal;
performance simulations show that with this method the coding
gain is almost entirely cancelled out (see Fig. 3).
Suboptimal decoding also provides a good initial bound of
the distance of the received point from the lattice, which can
be used as a cost function for the stack decoder described in
Fig. 4. Comparison between suboptimal decoding and ML decoding for the
RS(6; 3; 4) code at 2 bpcu. The first method achieves a gain of 2.4 dB over the Section VII-B.
uncoded case, compared to the 7.0 dB of ML decoding. • 2 bpcu: Fig. 5 shows the performance comparison of the
Golden-RS codes with suboptimal decoding with the un-
coded scheme at the spectral efficiency of 2 bpcu.
The gain for the code is unexpectedly high com- Assuming the channel to be constant for 4, 8, and 12
pared with the theoretical coding gain (24) for , that is blocks, respectively, the , and
dB dB. The rough estimate (24) is based on Golden-RS codes outperform the uncoded scheme at the
the worst possible occurrence, that of a codeword of Hamming same spectral efficiency by 1.1, 1.7, and 2.8 dB at the FER
weight in which all three nonzero components correspond to of .
invertible elements in the quotient. The Golden-RS schemes seem to be more robust on
However, we can verify empirically that in the 4-QAM case slow fading channels; in fact the performances of the
and with our choice of the code, this event does not take Golden- codes on a channel which is con-
place and in fact the actual value for found by computer stant for blocks remain almost unchanged (the variation
search is , giving an estimate for the gain of 3.2 dB, a little is less than 0.2 dB) when varies between 4 and 12, while
closer to the observed value. the uncoded Golden Code has a loss of almost 1.5 dB.
This favorable behavior might be due to the fact that the • 3 bpcu: Assuming the channel to be constant for 8, 16,
chosen constellation contains only one point in each coset, so and 24 blocks, respectively, the , and
that the codewords of Hamming distance are few. Golden-RS codes gain 1.5, 2.2, and 2.8 dB over
Also for the code, the actual gain (7.0 dB) is higher the uncoded scheme at the FER of (see Fig. 6).
than the theoretical gain ( dB dB based solely Similarly to the previous case, the Golden-
on the minimum Hamming distance; 5.3 dB using the true value codes lose less than 0.3 dB when varies between 8 and
of , that is .) 24, while the Golden Code has a loss of 1.1 dB.
LUZZI et al.: GOLDEN SPACE–TIME BLOCK-CODED MODULATION 593

Fig. 7. The output of the Reed–Solomon code and the uncoded bits are “min-
gled” before modulation.

Fig. 6. Performance of (8; 6; 3), (16; 12; 5), and (24; 18; 7) Golden
Reed–Solomon codes with suboptimal decoding at 3 bpcu compared to the
uncoded Golden Code scheme with the same spectral efficiency.

E. The 16-QAM Case

Using 16-QAM modulation for each symbol , , , in a
Golden codeword, there are available Golden codewords,
or 256 words for each of the 256 cosets of in .
As in the 4-QAM case, we consider coset codes where
the outer code is an Reed–Solomon code on
Fig. 8. The labelling of the 16-QAM constellation used for performance sim-
the quotient . Intuitively, the minimum distance of the ulations. The first and second bit identify one of the four cosets of 2 [i] in [i]
Reed–Solomon code “protects” the cosets from being decoded (drawn in different shades of gray); the third and fourth bit identify one of the
wrongly; if this choice is correct, the estimate for the right point four points in the coset. We remark that this type of labeling cannot be a Gray
mapping.
in the coset is protected by the minimum determinant in .
The total information bits transmitted are ; they will
be encoded into bits. to use long codes with a high minimum distance with this
— The code outputs bits, which are used to encode scheme.
the first two bits of 16-QAM constellations, that is the — If , , making for a gain of 2.5
bits which identify one of the four cosets of in ; dB.
each byte corresponds to a different coset configuration of
(see Fig. 8). F. Decoding
— the other bits, left uncoded, are used to choose the last
two bits of each 16-QAM signal. The ML decoding procedure for the 16-QAM case requires
In total, we have 16-QAM symbols, that is a vector of only a slight modification with respect the 4-QAM case illus-
Golden codewords . The resulting spectral trated in Section VII-B. In the first phase, for each component
efficiency is and for each coset leader , , we
find the closest point in that coset to the received component ,
bits
bpcu that is
channel uses
In this case, the coding gain depends on the minimum determi-
nant of the ideal in addition to the minimum Hamming distance
in the quotient: we have seen in (22) that Computing and separately allows to perform only
512 products instead of . The second phase can be per-
formed as in the 4-QAM case, and the search is limited to the
“closest points” determined in the previous phase,
i.e.,

With an error-correcting code of rate , the spectral

efficiency is 6 bpcu.
— If , , leading to an approx- over all the images of Reed–Solomon code-
imate gain of 3.8 dB. Thus, it does not seem worthwhile words.
594 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 55, NO. 2, FEBRUARY 2009

We need to show that .

But , and therefore
, and

recalling that .
Proof of Remark 3:
a) Let
Fig. 9. Performance of the (4; 2; 3) and (6; 3; 4) Golden Reed–Solomon codes
with ML decoding at 6 bpcu compared to the uncoded schemes with the same
spectral efficiency.

where . Then
. But for some
G. Simulation Results , and
In the 16-QAM case, the and Golden
Reed–Solomon codes achieve a gain of 3.9 and 4.3 dB, respec-
tively, over the uncoded scheme at 6 bpcu at the frame error
rate of , supposing that the channel is constant for 4 and 6
time blocks (see Fig. 9).
The same is true for .
VIII. CONCLUSIONS AND PERSPECTIVES b) If , then
In this paper, we have presented Golden-RS codes, a coded
modulation scheme for slow-fading MIMO channels,
where the inner code is the Golden Code.
We consider a simple binary partitioning based on a two-sided and
ideal of the Golden Code, whose set of coset leaders coincides
with a QAM symbol constellation. With a Reed–Solomon code
as the outer code in order to increase the minimum Hamming c) Let , . Then
distance among the codewords, we obtain a significant perfor-
mance gain with respect to the uncoded case.
Future work will deal with exploiting the ring structure of the
quotient to improve the overall distribution of determinants, in-
stead of focusing only on the minimum determinant. Our coded
modulation approach could also be used to improve the perfor- since belongs to .
mance of the and Perfect Codes in [8] in the slow
fading case.
APPENDIX B
QUATERNION ALGEBRAS
APPENDIX A
This section summarizes some basic facts about quaternion
PROOFS
algebras that are used in the paper. Our main references are the
We report here some of the proofs for the results stated in the books of Vignéras [17] and Reiner [18].
main part of the paper.
Definition 1 (Quaternion Algebras): Let be a field. A
Proof of Lemma 1: For all , let :
quaternion algebra of center is a central simple algebra of
then
dimension over , such that there exists a separable quadratic
extension of , and an element , with
LUZZI et al.: GOLDEN SPACE–TIME BLOCK-CODED MODULATION 595

where is the non-trivial -automorphism of . is called • normal if and are maximal,

a maximal subfield of . will be denoted by the triple • integral if , ,
. • principal if for some
Quaternion algebras are a special case of cyclic algebras. To The inverse of is the fractional ideal
obtain a representation of as a -module, consider a prim- .
itive element such that , and let , The norm of an ideal is the set of reduced norms of
. Then its elements, and it is an ideal of . If is principal,
.
(25)

The following theorem, which follows from a more general re- APPENDIX C
sult about cyclic algebras ([18, Corollary 30.7]) gives a suffi- IDEALS, VALUATIONS AND MAXIMAL ORDERS
cient condition for a quaternion algebra to be a division ring. Definition 5 (Valuations and Local Fields): A valuation
of is a positive real function of such that ,
Theorem 7: Let be a quaternion algebra.
a) ,
If is not a reduced norm of any element of , then is a skew
b) ,
field.
c) .
Definition 2 (Splitting Fields): Let be a central simple is non-Archimedean if
-algebra. An extension field of splits , or is a splitting ; it is discrete if is an infinite cyclic group.
field for , if can be endowed with a topology induced by in the fol-
lowing way: a neighborhood basis of a point is given by the
sets
In the case of division algebras, every maximal subfield is a
splitting field ([18, Th. 7.15]):
Theorem 8: Let be a skew field with center , with finite will be called complete if it is complete with respect to this
degree over . Then every maximal subfield of contains topology.
, and is a splitting field for . If is non-Archimedean, the set

In the following paragraphs we will always consider a

Dedekind domain , its quotient field , and a quaternion
algebra over . is a local ring, called the valuation ring of . The quotient
, where is the unique maximal ideal of , is called
Definition 3 (Lattices and Orders): A full -lattice or ideal the field of residues of .
in is a finitely generated -submodule in such that is a local field if it is complete with respect to a discrete
, where valuation and if is finite.
Definition 6 (Places): A place of is an immersion
into a local field . If is non-archimedean,
we say that it is a finite place; otherwise, that it is an infinite
An -order in is a full -lattice which is also a subring of place.
with the same unity element. A maximal -order is an order
which is not properly contained in any other order of . The finite places of arise from discrete -adic valuations
The following proposition is a consequence of [18, Th. 10.3]. of , where ranges over the maximal ideals in the ring of
integers of . (Recall that the ring of integers in a number
Proposition 9: A subring of containing a basis for field is always a Dedekind domain, and so the maximal ideals
over is an order if and only if all its elements are integral coincide with the prime ideals).
over . In the case of infinite places , the -adic completion can be
Remark 6: The notion of order is a generalization of the (real primes) or (complex primes).
notion of the ring of integers for commutative extensions. How- The notion of ramification for quaternion algebras is a gener-
ever, in the noncommutative case the set of elements which are alization of the notion of ramification for field extensions.
integral over the base field might not be a ring. Definition 7 (Ramified Places): Let be a quaternion al-
Definition 4 (Properties of Ideals): Given an ideal of , gebra over , and a place of .
we can define the left order and the right order of as follows: Consider the -module ; is isomorphic
to a matrix algebra over a skew field of center
and index over ; is called the local index of at .
We say that is ramified in if .
Complex primes are never ramified [18].
and are orders. is called Given a maximal order , the set of ramified places
• two-sided if , of is related to a particular two-sided ideal of :
596 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 55, NO. 2, FEBRUARY 2009

Definition 8 (Different and Discriminant): Let be an If is prime, from (26) we have . Since two-sided
order. The set ideals can be decomposed into ideals of and prime two-sided
ideals, they are invariant under involution.

REFERENCES
is a two-sided ideal, called the dual of . Its inverse [1] V. Tarokh, N. Seshadri, and A. R. Calderbank, “Space-time codes for
is a two-sided integral ideal, called the different of . high data rate wireless communication: Performance criterion and code
construction,” IEEE Trans. Inf. Theory, vol. 44, pp. 744–765, 1998.
If is a basis of as a free -module, [2] J.-C. Guey, M. P. Fitz, M. R. Bell, and W.-Y. Kuo, “Signal design for
transmitter diversity wireless communication systems over Rayleigh
fading channels,” in Proc. IEEE VTC ’96, pp. 136–140.
[3] C. Hollanti and J. Lahtonen, “A new tool: Constructing STBCs from
maximal orders in central simple algebras,” in Proc. Inf. Theory Work-
The ideal of is called the reduced discriminant of shop, 2006, pp. 322–326.
and is denoted by . [4] C. Hollanti, J. Lahtonen, and H. F. Lu, “Maximal orders in the design
of dense space-time lattice codes,” IEEE Trans. Inf. Theory, vol. 54,
Proposition 10: If , are two orders and , then no. 10, pp. 4493–4510, Oct. 2008.
. [5] C. Hollanti, J. Lahtonen, K. Ranto, and R. Vehkalahti, “On the densest
MIMO lattices from cyclic division algebras,” IEEE Trans. Inf. Theory,
For the following results see [17, Corollaire 5.3 and p. 86], Dec. 2006, submitted for publication.
respectively. [6] B. A. Sethuraman, B. S. Rajan, and V. Shashidar, “Full-diversity, high-
rate space-time block codes from division algebras,” IEEE Trans. Inf.
Proposition 11: Let be a quaternion algebra unramified Theory, vol. 49, pp. 2596–2616, 2003.
at infinity. [7] J.-C. Belfiore, G. Rekaya, and E. Viterbo, “The Golden Code: A 2 2 2
full-rate space-time code with non-vanishing determinants,” IEEE
A necessary and sufficient condition for an order to be max- Trans. Inf. Theory, vol. 51, 2005.
imal is that [8] F. Oggier, G. Rekaya, J.-C. Belfiore, and E. Viterbo, “Perfect space-
time blocks codes,” IEEE Trans. Inf. Theory, vol. 52, 2006.
[9] P. Elia, B. A. Sethuraman, and P. V. Kumar, “Perfect space-time codes
with minimum and non-minimum delay for any number of transmit
antennas,” IEEE Trans. Inf. Theory, vol. 53, pp. 3853–3868, 2007.
[10] S. Benedetto and E. Biglieri, Principles of Digital Transmission With
Definition 9 (Prime Ideals): Let be an order, a two- Wireless Applications. New York: Kluwer, 1999.
sided ideal of . is prime if it is nonzero and integer [11] P. Elia, K. R. Kumar, S. A. Pawar, P. V. Kumar, and H.-F. Lu, “Explicit,
two-sided ideals of , or . minimum delay space-time codes achieving the diversity-multiplexing
gain tradeoff,” IEEE Trans. Inf. Theory, vol. 52, pp. 3869–3884, 2006.
The following theorems are a consequence of more general [12] K. R. Kumar and G. Caire, “Space-time codes from structured lattices,”
results for central simple algebras ([18, Ths. 22.3, 22.4, 22.10, IEEE Trans. Inf. Theory, Apr. 2008, submitted for publication.
25.7]). [13] Y. Hong, E. Viterbo, and J.-C. Belfiore, “Golden space-time Trellis
coded modulation,” IEEE Trans. Inf. Theory, vol. 53, 2007.
Theorem 12: The two-sided ideals of an order form a [14] D. Champion, J.-C. Belfiore, G. Rekaya, and E. Viterbo, “Partitionning
the Golden Code: A framework to the design of space-time coded mod-
free group generated by the prime ideals. ulation,” presented at the Canadian Workshop on Information Theory,
2005.
Theorem 13: Let be a maximal order in a quaternion al- [15] T. W. Hungerford, Algebra. New York: Springer-Verlag, 1974.
gebra . Then the prime ideals of coincide with the maximal [16] G. D. Forney, “Coset codes- Part I: Introduction and geometrical clas-
two-sided ideals of , and there is a one-to-one correspondence sification,” IEEE Trans. Inf. Theory, vol. 34, 1988.
[17] M.-F. Vignéras, “Arithmétique des algèbres de quaternions,” in Lecture
between the prime ideals in and the prime ideals of , Notes in Mathematics. New York: Springer Verlag, 1980.
given by . [18] I. Reiner, Maximal Orders. Oxford, U.K.: Clarendon Press, 2003.
Moreover, is a simple algebra over the finite field .
Theorem 14: Let be a maximal order in . For each
place of , let be the local index of at , and let
be the prime ideal of corresponding to (see Theorem 13).
Then only for a finite number of places , and

(26)

Theorem 15: The two-sided ideals of a maximal order

form a commutative group with respect to multiplication, which
is generated by the ideals of and the ideals of reduced norm
, where varies over the prime ideals of that are ramified
in .
Remark 7: For any prime ideal of , let
. Since , and divide the same prime
.